tag:theconversation.com,2011:/au/topics/deepmind-25837/articlesDeepMind – The Conversation2024-03-19T19:44:12Ztag:theconversation.com,2011:article/2258942024-03-19T19:44:12Z2024-03-19T19:44:12ZCan AI improve football teams’ success from corner kicks? Liverpool and others are betting it can<figure><img src="https://images.theconversation.com/files/582686/original/file-20240318-26-ut2che.jpg?ixlib=rb-1.1.0&rect=0%2C0%2C4000%2C2005&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><span class="source">Google DeepMind</span></span></figcaption></figure><p>Last Sunday, Liverpool faced Manchester United in the <a href="https://www.espn.com.au/football/report/_/gameId/699283">quarter finals of the FA Cup</a> – and in the final minute of extra time, with the score tied at three-all, Liverpool had the crucial opportunity of a corner kick. A goal would surely mean victory, but losing possession could be risky.</p>
<p>What was Liverpool to do? Attack or play it safe? And if they were to attack, how best to do it? What kind of delivery, and where should players be waiting to attack the ball?</p>
<p>Set-piece decisions like this are vital not only in football but in many other competitive sports, and traditionally they are made by coaches on the basis of long experience and analysis. However, Liverpool has recently been looking to an unexpected source for advice: researchers at the Google-owned UK-based artificial intelligence (AI) lab <a href="https://deepmind.google/discover/blog/advancing-sports-analytics-through-ai-research/">DeepMind</a>.</p>
<p>In a <a href="https://www.nature.com/articles/s41467-024-45965-x">paper published today</a> in Nature Communications, DeepMind researchers describe an AI system for football tactics called TacticAI, which can assist in developing successful corner kick routines. The paper says experts at Liverpool favoured TacticAI’s advice over existing tactics in 90% of cases.</p>
<h2>What TacticAI can do</h2>
<p>At a corner kick, play stops and each team has the chance to organise its players on the field before the attacking team kicks the ball back into play – usually with a specific prearranged plan in mind that will (hopefully) let them score a goal. Advice on these prearranged plans or routines is what TacticAI sets out to offer.</p>
<p>The package has three components: one that predicts which player is most likely to receive the ball in a given scenario, another that predicts whether a shot on goal will be taken, and a third that recommends how to adjust the position of players to increase or decrease the chances of a shot on goal.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="A diagram showing a soccer field with player positions marked, as well as a network diagram." src="https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=258&fit=crop&dpr=1 600w, https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=258&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=258&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=325&fit=crop&dpr=1 754w, https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=325&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/582707/original/file-20240319-28-xag9u9.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=325&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">TacticAI represents a corner-kick setup as a ‘graph’ of player positions and relationships, which it then uses to make predictions.</span>
<span class="attribution"><a class="source" href="https://doi.org/10.1038/s41467-024-45965-x">Wang et al. / Nature Communications</a></span>
</figcaption>
</figure>
<p>Trained on a dataset of 7,176 corner kicks from Premier League matches, TacticAI used a technique called “geometric deep learning” to identify key strategic patterns.</p>
<p>The researchers say this approach could be applied not only to football, but to any sport in which a stoppage in the game allows teams to deliberately manoeuvre players into place unopposed, and plan the next sequence of play. In football, it could also be expanded in future to incorporate throw-in routines as well as other set pieces such as attacking free kicks.</p>
<h2>Vast amounts of data</h2>
<p>AI in football is not new. Even in amateur and semi-professional football, AI-powered auto-tracking camera systems are becoming commonplace, for example. At the last men’s and women’s World Cups in 2022 and 2023, AI in conjunction with advanced ball-tracking technology produced semi-automated offside decisions with an unprecedented level of accuracy.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/games-by-numbers-machine-learning-is-changing-sport-38973">Games by numbers: machine learning is changing sport</a>
</strong>
</em>
</p>
<hr>
<p>Professional football clubs have analytical departments using AI at every level of the game, predominantly in the areas of <a href="https://www.wired.com/story/ai-football-soccer-scouting/">scouting</a>, <a href="https://www.engadget.com/will-ai-revolutionize-professional-soccer-recruitment-130045118.html">recruitment</a> and <a href="https://theathletic.com/4966509/2023/10/19/wearable-technology-in-football/">athlete monitoring</a>. Other research has also tried to <a href="https://www.mdpi.com/1424-8220/23/9/4506">predict players’ shots on goal</a>, or guess from a video what <a href="https://www.nature.com/articles/s41598-022-12547-0">off-screen players are doing</a>. </p>
<p>Bringing AI into tactical decisions promises to offer coaches a more objective and analytical approach to the game. Algorithms can process vast amounts of data, identifying patterns that may not be apparent to the naked eye, giving teams valuable insights into their own performance as well as that of their opponents. </p>
<h2>A useful tool</h2>
<p>AI may be a useful tool, but it cannot make decisions about match play alone. An algorithm might suggest the optimal positional setup for an in-swinging corner or how best to exploit the opposition’s defensive tactics. </p>
<p>What AI cannot do is make decisions on the fly – like deciding whether to take a corner quickly to exploit an opponent’s lapse in concentration. </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/b1zjjf5EN1g?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">Sometimes the best move is a speedy reaction to conditions on the ground, not an elaborate prearranged set play.</span></figcaption>
</figure>
<p>There’s also something to be said for allowing players creative licence in some situations. Once teams are using AI to suggest the optimal corner strategy, opponents will doubtless counter with their own AI-prompted defensive setup.</p>
<p>So while the tech behind TacticAI is very interesting, it remains to be seen whether it can evolve to be useful in open play. Could AI get to the stage where it can recognise the best tactical player substitution in a given situation? </p>
<p>DeepMind researchers have advanced decision-making like this in their sights for <a href="https://dl.acm.org/doi/10.1613/jair.1.12505">future research</a>, but will it ever reach a point where coaches would trust it?</p>
<p>My sense from discussions with people in the industry is many believe AI should only be used as an input to decision-making, and not be allowed to make decisions itself. There is no substitute for the experience and instinct of the best coaches, the intangible ability to feel what the game needs, to make a change in formation, to play someone out of position. </p>
<h2>Smart tactics – but what about strategy?</h2>
<p>Coming back to that crucial Liverpool corner in last Sunday’s FA Cup quarter final: we don’t know whether Liverpool’s manager Jürgen Klopp considered AI advice, but the decision was made to play an attacking corner kick, presumably in the hope of scoring a last-minute winner. </p>
<p>The out-swinging delivery into the box may well have been the tactic with the highest probability of scoring a goal – but things rapidly went wrong. Manchester United gained possession of the ball, moved it down the pitch on the counterattack and slotted home the winning goal, sending Liverpool out of the tournament at the last moment.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/DKk8N2PYwCA?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">Even the best tactics can go wrong.</span></figcaption>
</figure>
<p>So while AI might suggest the optimal delivery and setup for a set piece, a coach might decide the wiser move is to play safe and avoid the risk of a counterattack. If TacticAI continues its career progression as a coaching assistant, it will no doubt learn that keeping the ball in the corner and playing for penalties may sometimes be the better option.</p><img src="https://counter.theconversation.com/content/225894/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Mark Scanlan does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>A new AI system may improve soccer tactics in 90% of corner kicks – but is it ready for the big leagues?Mark Scanlan, Lecturer, Edith Cowan UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/2066142023-06-01T00:37:14Z2023-06-01T00:37:14ZNo, AI probably won’t kill us all – and there’s more to this fear campaign than meets the eye<figure><img src="https://images.theconversation.com/files/529293/original/file-20230531-27-xt3dun.jpeg?ixlib=rb-1.1.0&rect=0%2C0%2C2525%2C1402&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><a class="source" href="https://betterimagesofai.org/images?artist=AlanWarburton&title=SocialMedia">Better Images of AI / Alan Warburton</a>, <a class="license" href="http://creativecommons.org/licenses/by-sa/4.0/">CC BY-SA</a></span></figcaption></figure><p>Doomsaying is an old occupation. Artificial intelligence (AI) is a complex subject. It’s easy to fear what you don’t understand. These three truths go some way towards explaining the oversimplification and dramatisation plaguing discussions about AI. </p>
<p>Yesterday outlets around the world were plastered with news of yet another <a href="https://www.safe.ai/statement-on-ai-risk">open letter claiming</a> AI poses an existential threat to humankind. This letter, published through the nonprofit Center for AI Safety, has been signed by industry figureheads including <a href="https://theconversation.com/ai-pioneer-geoffrey-hinton-says-ai-is-a-new-form-of-intelligence-unlike-our-own-have-we-been-getting-it-wrong-this-whole-time-204911">Geoffrey Hinton</a> and the chief executives of Google DeepMind, Open AI and Anthropic. </p>
<p>However, I’d argue a healthy dose of scepticism is warranted when considering the AI doomsayer narrative. Upon close inspection, we see there are commercial incentives to manufacture fear in the AI space. </p>
<p>And as a researcher of artificial general intelligence (AGI), it seems to me the framing of AI as an existential threat has more in common with 17th-century philosophy than computer science.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/ai-pioneer-geoffrey-hinton-says-ai-is-a-new-form-of-intelligence-unlike-our-own-have-we-been-getting-it-wrong-this-whole-time-204911">AI pioneer Geoffrey Hinton says AI is a new form of intelligence unlike our own. Have we been getting it wrong this whole time?</a>
</strong>
</em>
</p>
<hr>
<h2>Was ChatGPT a ‘breakthrough’?</h2>
<p>When ChatGPT was released late last year, people were delighted, entertained and horrified. </p>
<p>But ChatGPT isn’t a research breakthrough as much as it is a product. The technology it is based on is several years old. An early version of its underlying model, GPT-3, was released in 2020 with many of the same capabilities. It just wasn’t easily accessible online for everyone to play with.</p>
<p>Back in 2020 and 2021, <a href="https://ieeexplore.ieee.org/document/9495946">I</a> and many <a href="https://link.springer.com/article/10.1007/s11023-020-09548-1">others</a> wrote papers discussing the capabilities and shortcomings of GPT-3 and similar models – and the world carried on as always. Forward to today, and ChatGPT has had an incredible impact on society. What changed?</p>
<p>In March, Microsoft researchers <a href="https://futurism.com/gpt-4-sparks-of-agi">published a paper</a> claiming GPT-4 showed “sparks of artificial general intelligence”. AGI is the subject of a variety of competing definitions, but for the sake of simplicity can be understood as AI with human-level intelligence.</p>
<p>Some immediately interpreted the Microsoft research as saying GPT-4 <em>is</em> an AGI. By the definitions of AGI I’m familiar with, this is certainly not true. Nonetheless, it added to the hype and furore, and it was hard not to get caught up in the panic. Scientists are no more immune to <a href="https://link.springer.com/book/10.1007/978-3-030-36822-7">group think</a> than anyone else.</p>
<p>The same day that paper was submitted, The Future of Life Institute <a href="https://futureoflife.org/open-letter/pause-giant-ai-experiments/">published an open letter</a> calling for a six-month pause on training AI models more powerful than GPT-4, to allow everyone to take stock and plan ahead. Some of the AI luminaries who signed it expressed concern that AGI poses an existential threat to humans, and that ChatGPT is too close to AGI for comfort. </p>
<p>Soon after, prominent AI safety researcher Eliezer Yudkowsky – who has been commenting on the dangers of superintelligent AI <a href="https://intelligence.org/files/AIPosNegFactor.pdf">since well before</a> 2020 – took things a step further. <a href="https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/">He claimed</a> we were on a path to building a “superhumanly smart AI”, in which case “the obvious thing that would happen” is “literally everyone on Earth will die”. He even suggested countries need to be willing to risk nuclear war to enforce compliance with AI regulation across borders. </p>
<h2>I don’t consider AI an imminent existential threat</h2>
<p>One aspect of AI safety research is to address potential dangers AGI might present. It’s a difficult topic to study because there is little agreement on what intelligence is and how it functions, let alone what a superintelligence might entail. As such, researchers must rely as much on speculation and philosophical argument as on evidence and mathematical proof.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/has-gpt-4-really-passed-the-startling-threshold-of-human-level-artificial-intelligence-well-it-depends-202856">Has GPT-4 really passed the startling threshold of human-level artificial intelligence? Well, it depends</a>
</strong>
</em>
</p>
<hr>
<p>There are two reasons I’m not concerned by ChatGPT and its <a href="https://lablab.ai/blog/what-is-babyagi-and-how-can-i-benefit-from-it">byproducts</a>. </p>
<p>First, it isn’t even close to the sort of artificial superintelligence that might conceivably pose a threat to humankind. The models underpinning it are slow learners that require immense volumes of data to construct anything akin to the versatile concepts humans can concoct from only a few examples. In this sense, it is not “intelligent”.</p>
<p>Second, many of the more catastrophic AGI scenarios depend on premises I find implausible. For instance, there seems to be a prevailing (but unspoken) assumption that sufficient intelligence amounts to limitless real-world power. If this was true, more scientists would be billionaires. </p>
<p>Moreover, cognition as we understand it in humans takes place as part of a physical environment (which includes our bodies), and this environment imposes limitations. The concept of AI as a “software mind” unconstrained by hardware has more in common with 17th-century <a href="https://plato.stanford.edu/entries/dualism/">dualism</a> (the idea that the mind and body are separable) than with contemporary theories of the mind existing as <a href="https://plato.stanford.edu/entries/embodied-cognition/">part of the physical world</a>. </p>
<h2>Why the sudden concern?</h2>
<p>Still, doomsaying is old hat, and the events of the last few years probably haven’t helped – but there may be more to this story than meets the eye. </p>
<p>Among the prominent figures calling for AI regulation, many work for or have ties to incumbent AI companies. This technology is useful, and there is money and power at stake – so fearmongering presents an opportunity.</p>
<p>Almost everything involved in building ChatGPT has been published in research anyone can access. OpenAI’s competitors can (and have) replicated the process, and it won’t be long before free and open-source alternatives flood the market.</p>
<p>This point was made clearly in a memo <a href="https://www.semianalysis.com/p/google-we-have-no-moat-and-neither">purportedly leaked</a> from Google entitled “We have no moat, and neither does OpenAI”. A moat is jargon for a way to secure your business against competitors.</p>
<p>Yann LeCun, who leads AI research at Meta, says these models should be open since they will become public infrastructure. He and many others are <a href="https://www.businesstoday.in/technology/news/story/completely-ridiculous-metas-chief-ai-scientist-yann-lecun-dismisses-elon-musks-civilisation-destruction-fear-383371-2023-05-30">unconvinced by the AGI doom</a> narrative. </p>
<p><div data-react-class="Tweet" data-react-props="{"tweetId":"1659172655663030272"}"></div></p>
<p>Notably, <a href="https://fortune.com/2023/05/05/meta-mark-zuckerberg-not-invited-ai-meeting-white-house/">Meta wasn’t invited</a> when US President Joe Biden recently met with the leadership of Google DeepMind and OpenAI. That’s despite the fact that Meta is almost certainly a leader in AI research; it produced PyTorch, the machine-learning framework OpenAI used to make GPT-3.</p>
<p>At the White House meetings, OpenAI chief executive Sam Altman suggested the US government should issue licences to those who are trusted to responsibly train AI models. Licences, as Stability AI chief executive Emad Mostaque <a href="https://twitter.com/EMostaque/status/1658653142429450242?s=20">puts it</a>, “are a kinda moat”. </p>
<p>Companies such as Google, OpenAI and Microsoft have everything to lose by allowing small, independent competitors to flourish. Bringing in licensing and regulation would help cement their position as market leaders and hamstring competition before it can emerge. </p>
<p>While regulation is appropriate in some circumstances, regulations that are rushed through will favour incumbents and suffocate small, <a href="https://www.forbes.com/sites/hessiejones/2023/04/19/amid-growing-call-to-pause-ai-research-laion-petitions-governments-to-keep-agi-research-open-active-and-responsible/?sh=1b21161a62e3">free and open-source competition</a>.</p>
<p><div data-react-class="Tweet" data-react-props="{"tweetId":"1654357631514079233"}"></div></p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/calls-to-regulate-ai-are-growing-louder-but-how-exactly-do-you-regulate-a-technology-like-this-203050">Calls to regulate AI are growing louder. But how exactly do you regulate a technology like this?</a>
</strong>
</em>
</p>
<hr>
<img src="https://counter.theconversation.com/content/206614/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Michael Timothy Bennett does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>I study artificial general intelligence, and I believe the ongoing fearmongering is at least partially attributable to large AI developers’ financial interests.Michael Timothy Bennett, PhD Student, School of Computing, Australian National UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1769522023-01-24T06:16:06Z2023-01-24T06:16:06ZHow we cracked the mystery of Australia’s prehistoric giant eggs<figure><img src="https://images.theconversation.com/files/505414/original/file-20230119-20-rbmop2.jpg?ixlib=rb-1.1.0&rect=0%2C0%2C1488%2C927&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">The giant bird Genyornis went extinct in Australia around 50,000 years ago.</span> <span class="attribution"><span class="source">Peter Trusler</span>, <span class="license">Author provided</span></span></figcaption></figure><p>It’s a long-running Australian detective story. From the 1980s onwards, researchers found eggshell fragments, and on rare occasions whole eggs, exposed in eroding sand dunes within the country’s arid zone (which covers most of Australia’s landmass).</p>
<p>A proportion of shells matched eggs laid by emus, but the rest belonged to a mystery species. Researchers initially identified the eggshells as belonging to a giant, extinct bird called <em>Genyornis</em>. But more recently, a group of scientists challenged this view.</p>
<p>With the help of artificial intelligence software, our team has now resolved this scientific controversy, showing that <em>Genyornis</em> was indeed the bird that laid these eggs. With colleagues based around the world, we have published the findings in <a href="https://www.pnas.org/doi/full/10.1073/pnas.2109326119">Proceedings of the National Academy of Sciences</a>. </p>
<p><em>Genyornis</em> was a flightless bird between two metres and 2.5 metres tall that once roamed the Australian landmass. The eggshell fragments are an important line of evidence about this extinct creature, so being certain about the identity of the bird that laid them is vital.</p>
<p>Some of the shell fragments are 400,000 years old, while the youngest are about 50,000 years old. <a href="https://www.nature.com/articles/ncomms10496">Previous work</a> showed that some of the youngest eggshells had been burned, but not in the way a wildfire would. Instead, scientific tests point to humans cooking the eggs for food.</p>
<p>The time period where <em>Genyornis</em> shells disappear (50,000 years ago) coincides with what’s thought to be the first arrival of humans in Australia. The discovery therefore raises the possibility that our species contributed to its extinction.</p>
<h2>Narrowing the candidates</h2>
<p>The eggshell fragments were first recognised by Dom Williams, a geologist and vertebrate palaeontologist from Flinders University in Adelaide, in 1981. He made the case that the fragments <a href="https://www.tandfonline.com/doi/abs/10.1080/03115518108565426">came from <em>Genyornis</em></a>, which belonged to a group of extinct creatures known as thunderbirds.</p>
<p>In the 1990s, a team including John Magee, at Australian National University, and Gifford Miller, one of the authors of this article, <a href="https://www.science.org/doi/abs/10.1126/science.283.5399.205">provided firm dates</a> for similar shell fragments collected at thousands of arid zone sites. <em>Genyornis</em> was one of many large animals – known as “megafauna” – that once roamed Australia and vanished at around the same time. The work by Miller, Magee and others pinned a clear date of 50,000 years ago on this extinction event.</p>
<p>The association of the eggshells with <em>Genyornis</em> was widely accepted from the 1980s until recently, when it was <a href="https://www.sciencedirect.com/science/article/pii/S027737911530192X">challenged by a team of scientists</a> from Flinders University in Australia. Based upon the size and structure of the eggshells, they argued for a different parent. Their favoured candidate was <em>Progura</em>, a 10kg extinct relative of modern birds such as the brush turkey and malleefowl. </p>
<p>Living birds belonging to this group - known as megapodes – build earthen mounds to incubate their eggs. <a href="https://theconversation.com/a-case-of-mistaken-identity-for-australias-extinct-big-bird-52856">The scientific debate</a> was fought out in academic journals, with neither side conceding.</p>
<h2>Chasing a solution</h2>
<p>Attempting to find a resolution, scientists who thought the eggs belonged to Genyornis turned to DNA. Despite <a href="https://royalsocietypublishing.org/doi/full/10.1098/rspb.2009.2019">the successful extraction</a> of genetic information from eggs of New Zealand’s extinct Moa bird, state-of-the-art DNA sequencing technology drew a blank in this case. The molecules were too degraded after 50,000 years under the hot Australian sun.</p>
<p>However, proteins – the molecular building blocks of cells – can provide similar information and can last for longer than DNA. In our study, we used a technique called amino acid racemisation to identify the shell fragments with the best-preserved proteins.</p>
<p>As part of the work, our team was able to retrieve partial protein sequences from the Australian eggshells. We then used software called AlphaFold, from the Google-owned AI lab DeepMind, to generate predicted structures for the molecules – the first time this has been done for ancient proteins.</p>
<p>Two of us, Matthew Collins and Beatrice Demarchi, contacted the <a href="https://b10k.genomics.cn">Bird 10,000 Genomes (B10K) Project</a>. This has set itself the ambitious goal of sequencing the genomes of all bird species.</p>
<p>B10K project member Josefin Stiller took the reconstructed protein sequences and <a href="https://unfolded.deepmind.com/stories/unlocking-the-mystery-of-the-demon-duck-of-doom">placed them within a “family tree”</a> showing how proteins differ between bird species. The proteins were complete enough to resolve the position of the mystery eggs within the deep branches of this tree of protein sequences, but not sufficiently diagnostic to uniquely identify what the parent bird was.</p>
<p>However, as <a href="https://www.pnas.org/doi/full/10.1073/pnas.2109326119">detailed in our latest paper</a>, the protein sequences were able to conclusively rule out that the parent was a megapode. As there are no other candidate birds, we concluded – as Williams had first proposed in the 1980s – that the eggshells belonged to <em>Genyornis</em>.</p>
<p>This means we can confidently interpret other evidence locked in the shells with implications for how <em>Genyornis</em> went extinct and why the emus that lived alongside it survived. </p>
<h2>Picky eater</h2>
<p>Isotopes are different forms of chemical elements that can record information about factors such as diet and climate. Carbon isotopes within the eggshell fragments provide information on the birds’ diets and show that <em>Genyornis</em> was a pickier eater than the emu. Oxygen isotopes can be used to track aridity and show that conditions were increasingly dry around the time <em>Genyornis</em> eggshells disappear.</p>
<p>In previous work, Miller and his colleagues <a href="https://www.sciencedirect.com/science/article/abs/pii/S0277379116302815">analysed the same isotopes in emu eggshells</a> across the time window of <em>Genyornis’</em> extinction and found that summer-season grasses abruptly disappear from the birds’ diets. This is consistent with a dramatic reduction in monsoon rains.</p>
<p>These findings suggest that <em>Genyornis</em> was already somewhat vulnerable to a changing environment, but another factor may have proved important to its ultimate fate. </p>
<p>When coupled with the lack of evidence from <em>Genyornis</em> skeletons for direct predation, the burnt eggshells suggest that – as is so common elsewhere in the world – human pressure was likely to have been a factor that finally drove these impressive birds to extinction.</p><img src="https://counter.theconversation.com/content/176952/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Matthew James Collins receives funding from The Danish National Research Foundation. He is affiliated with The University of Copenhagen. </span></em></p><p class="fine-print"><em><span>Beatrice Demarchi receives funding from the Italian Ministry of University and Research </span></em></p><p class="fine-print"><em><span>Gifford Miller receives funding from the US National Science Foundation. </span></em></p>A puzzle over the identity of an extinct bird that laid eggs across Australia has been solved.Matthew James Collins, Professor of Palaeoproteomics, University of CambridgeBeatrice Demarchi, Associate professor, Università di TorinoGifford Miller, Distinguished Professor of Geological Sciences, University of Colorado BoulderLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1820312022-06-16T18:39:25Z2022-06-16T18:39:25ZA celebrated AI has learned a new trick: How to do chemistry<figure><img src="https://images.theconversation.com/files/469287/original/file-20220616-12-dmwhkp.jpg?ixlib=rb-1.1.0&rect=2%2C0%2C1794%2C840&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Figuring out what makes some proteins glow requires an understanding of chemistry.</span> <span class="attribution"><a class="source" href="https://www.flickr.com/photos/128643624@N07/16652974221/">eLife - the journal</a>, <a class="license" href="http://creativecommons.org/licenses/by-sa/4.0/">CC BY-SA</a></span></figcaption></figure><p>Artificial intelligence has changed the way science is done by allowing researchers to analyze the massive amounts of data modern scientific instruments generate. It can find a needle in a million haystacks of information and, using <a href="https://www.techtarget.com/searchenterpriseai/definition/deep-learning-deep-neural-network">deep learning</a>, it can learn from the data itself. AI is accelerating advances in <a href="https://doi.org/10.1186/s13073-021-00965-0">gene hunting</a>, <a href="https://doi.org/10.1038/s41591-020-01197-2">medicine</a>, <a href="https://news.mit.edu/2021/drug-discovery-binding-affinity-0315">drug design</a> and <a href="https://doi.org/10.1038/nature25978">the creation of organic compounds</a>.</p>
<p>Deep learning uses algorithms, often neural networks that are trained on large amounts of data, to extract information from new data. It is very different from traditional computing with its step-by-step instructions. Rather, it learns from data. Deep learning is far less transparent than traditional computer programming, leaving important questions – what has the system learned, what does it know?</p>
<p>As a <a href="https://scholar.google.ca/citations?user=RpiSPiwAAAAJ&hl=en">chemistry professor</a> I like to design tests that have at least one difficult question that stretches the students’ knowledge to establish whether they can combine different ideas and synthesize new ideas and concepts. We have devised such a question for the poster child of AI advocates, AlphaFold, which has solved the <a href="https://doi.org/10.1146%2Fannurev.biophys.37.092707.153558">protein-folding problem</a>.</p>
<h2>Protein folding</h2>
<p>Proteins are present in all living organisms. They provide the cells with structure, catalyze reactions, transport small molecules, digest food and do much more. They are made up of long chains of amino acids like beads on a string. But for a protein to do its job in the cell, it must twist and bend into a complex three-dimensional structure, a process called protein folding. Misfolded proteins can lead to disease.</p>
<p>In his chemistry Nobel acceptance speech in 1972, <a href="https://www.nobelprize.org/prizes/chemistry/1972/anfinsen/biographical/">Christiaan Anfinsen</a> postulated that it should be possible to <a href="https://directorsblog.nih.gov/tag/christian-anfinsen/">calculate the three-dimensional structure of a protein from the sequence of its building blocks</a>, the amino acids. </p>
<p>Just as the order and spacing of the letters in this article give it sense and message, so the order of the amino acids determines the protein’s identity and shape, which results in its function. </p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="a graphic showing a thread-like line on the left and a coiled structure on the right" src="https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=266&fit=crop&dpr=1 600w, https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=266&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=266&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=334&fit=crop&dpr=1 754w, https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=334&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/462041/original/file-20220509-23-mkr8t2.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=334&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Within milliseconds of the exit of an amino acid chain (left) from the ribosome, it is folded into the lowest-energy 3D shape (right), which is required for the protein’s function.</span>
<span class="attribution"><span class="source">Marc Zimmer</span>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span>
</figcaption>
</figure>
<p>Because of the inherent flexibility of the amino acid building blocks, a typical protein can adopt an estimated <a href="https://web.archive.org/web/20110523080407/http:/www-miller.ch.cam.ac.uk/levinthal/levinthal.html">10 to the power of 300 different forms</a>. This is a massive number, more than the <a href="https://educationblog.oup.com/secondary/maths/numbers-of-atoms-in-the-universe">number of atoms in the universe</a>. Yet within a millisecond every protein in an organism will fold into its very own specific shape – the lowest-energy arrangement of all the chemical bonds that make up the protein. Change just one amino acid in the hundreds of amino acids typically found in a protein and it may misfold and no longer work. </p>
<h2>AlphaFold</h2>
<p>For 50 years computer scientists have tried to solve the protein-folding problem – with little success. Then in 2016 <a href="https://www.deepmind.com/">DeepMind</a>, an AI subsidiary of Google parent Alphabet, initiated its <a href="https://www.deepmind.com/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology">AlphaFold</a> program. It used the <a href="https://www.rcsb.org/">protein databank</a> as its training set, which contains the experimentally determined structures of over 150,000 proteins. </p>
<p>In less than five years AlphaFold had <a href="https://www.deepmind.com/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology">the protein-folding problem beat</a> – at least the most useful part of it, namely, determining the protein structure from its amino acid sequence. AlphaFold does not explain how the proteins fold so quickly and accurately. It was a major win for AI, because it not only accrued huge scientific prestige, it also was a major scientific advance that could affect everyone’s lives.</p>
<p>Today, thanks to programs like <a href="https://www.deepmind.com/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology">AlphaFold2</a> and <a href="https://www.ipd.uw.edu/2021/07/rosettafold-accurate-protein-structure-prediction-accessible-to-all/">RoseTTAFold</a>, researchers like me can determine the three-dimensional structure of proteins from the sequence of amino acids that make up the protein – at no cost – in an hour or two. Before AlphaFold2 we had to crystallize the proteins and solve the structures using <a href="https://doi.org/10.1136%2Fmp.53.1.8">X-ray crystallography</a>, a process that took months and cost tens of thousands of dollars per structure. </p>
<p>We now also have access to the <a href="https://alphafold.ebi.ac.uk/">AlphaFold Protein Structure Database</a>, where Deepmind has deposited the 3D structures of nearly all the proteins found in humans, mice and more than 20 other species. To date they it has solved more than a million structures and plan to add another 100 million structures this year alone. Knowledge of proteins has skyrocketed. The structure of half of all known proteins is likely to be documented by the end of 2022, among them many new unique structures associated with new useful functions.</p>
<h2>Thinking like a chemist</h2>
<p>AlphaFold2 was not designed to predict how proteins would interact with one another, yet it has been able to model how individual proteins combine to <a href="https://www.nature.com/articles/d41586-022-00997-5">form large complex units composed of multiple proteins</a>. We had a challenging question for AlphaFold – had its structural training set taught it some chemistry? Could it tell whether amino acids would react with one another – a rare yet important occurrence?</p>
<p>I am a computational chemist interested in <a href="https://theconversation.com/fluorescent-proteins-light-up-science-by-making-the-invisible-visible-39272">fluorescent proteins</a>. These are proteins found in hundreds of marine organisms like jellyfish and coral. Their glow can be used <a href="https://theconversation.com/from-crispr-to-glowing-proteins-to-optogenetics-scientists-most-powerful-technologies-have-been-borrowed-from-nature-164459">to illuminate</a> and <a href="https://global.oup.com/academic/product/illuminating-disease-9780199362813?cc=us&lang=en&">study diseases</a>.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="two multicolored blobs with bright lines inside them against a black background" src="https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=302&fit=crop&dpr=1 600w, https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=302&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=302&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=379&fit=crop&dpr=1 754w, https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=379&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/468852/original/file-20220614-12-84y9j5.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=379&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Neurons expressing fluorescent proteins reveal the brain structures of two fruit fly larvae.</span>
<span class="attribution"><a class="source" href="https://images.nigms.nih.gov/pages/DetailPage.aspx?imageid2=6808">Wen Lu and Vladimir I. Gelfand, Feinberg School of Medicine, Northwestern University</a></span>
</figcaption>
</figure>
<p>There are 578 fluorescent proteins in the <a href="https://www.rcsb.org/search?request=%7B%22query%22%3A%7B%22type%22%3A%22group%22%2C%22nodes%22%3A%5B%7B%22type%22%3A%22group%22%2C%22nodes%22%3A%5B%7B%22type%22%3A%22group%22%2C%22nodes%22%3A%5B%7B%22type%22%3A%22terminal%22%2C%22service%22%3A%22text%22%2C%22parameters%22%3A%7B%22attribute%22%3A%22struct_keywords.pdbx_keywords%22%2C%22operator%22%3A%22contains_phrase%22%2C%22value%22%3A%22FLUORESCENT%20PROTEIN%22%7D%7D%5D%2C%22logical_operator%22%3A%22and%22%7D%5D%2C%22logical_operator%22%3A%22and%22%2C%22label%22%3A%22text%22%7D%5D%2C%22logical_operator%22%3A%22and%22%7D%2C%22return_type%22%3A%22entry%22%2C%22request_options%22%3A%7B%22paginate%22%3A%7B%22start%22%3A0%2C%22rows%22%3A25%7D%2C%22scoring_strategy%22%3A%22combined%22%2C%22sort%22%3A%5B%7B%22sort_by%22%3A%22score%22%2C%22direction%22%3A%22desc%22%7D%5D%7D%2C%22request_info%22%3A%7B%22query_id%22%3A%223e70236cf383b26f27688c5c79c6eb2b%22%7D%7D">protein databank</a>, of which 10 are “broken” and don’t fluoresce. Proteins rarely attack themselves, a process called autocatalytic posttranslation modification, and it is very difficult to predict which proteins will react with themselves and which ones won’t. </p>
<p>Only a chemist with a significant amount of fluorescent protein knowledge would be able to use the amino acid sequence to find the fluorescent proteins that have the right amino acid sequence to undergo the chemical transformations required to make them fluorescent. When we presented AlphaFold2 with the sequences of 44 fluorescent proteins that are not in the protein databank, <a href="https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0267560">it folded the fixed fluorescent proteins differently from the broken ones</a>.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="a diagram showing a light bulb on the left and the stem only of a light bulb on the right" src="https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=333&fit=crop&dpr=1 600w, https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=333&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=333&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=419&fit=crop&dpr=1 754w, https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=419&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/462050/original/file-20220509-12-fxhj9p.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=419&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">AlphaFold2 can take the amino acid sequence of fluorescent proteins (letters at the top) and predict their 3D barrel shapes (middle). This isn’t surprising. What is totally unexpected is that it can also predict which fluorescent proteins are ‘broken’ and can’t fluoresce.</span>
<span class="attribution"><span class="source">Marc Zimmer</span>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span>
</figcaption>
</figure>
<p>The result stunned us: AlphaFold2 had learned some chemistry. It had figured out which amino acids in fluorescent proteins do the chemistry that makes them glow. We suspect that the protein databank training set and <a href="https://samfordubioinformatics.wordpress.com/bioinformatics-techniques/multiple-sequence-alignment/">multiple sequence alignments</a> enable AlphaFold2 to “think” like chemists and look for the amino acids required to react with one another to make the protein fluorescent. </p>
<p>A folding program learning some chemistry from its training set also has wider implications. By asking the right questions, what else can be gained from other deep learning algorithms? Could facial recognition algorithms find hidden markers for diseases? Could algorithms designed to predict spending patterns among consumers also find a propensity for minor theft or deception? And most important, is this capability – and <a href="https://www.technologyreview.com/2019/09/17/75427/open-ai-algorithms-learned-tool-use-and-cooperation-after-hide-and-seek-games/">similar leaps in ability</a> in other AI systems – desirable?</p><img src="https://counter.theconversation.com/content/182031/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Marc Zimmer does not work for, consult, own shares in or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>The AI AlphaFold can figure out the three-dimensional protein structure any string of amino acids will become. It has now exceeded its training by figuring out what makes some proteins glow.Marc Zimmer, Professor of Chemistry, Connecticut CollegeLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1729002021-12-01T17:10:14Z2021-12-01T17:10:14ZMathematical discoveries take intuition and creativity – and now a little help from AI<figure><img src="https://images.theconversation.com/files/434904/original/file-20211201-25-xkwy6s.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">shutterstock</span> </figcaption></figure><p>Research in mathematics is a deeply imaginative and intuitive process. This might come as a surprise for those who are still recovering from high-school algebra.</p>
<p>What does the world look like at the quantum scale? What shape would our universe take if we were as large as a galaxy? What would it be like to live in six or even 60 dimensions? These are the problems that mathematicians and physicists are grappling with every day.</p>
<p>To find the answers, mathematicians like me try to find patterns that relate complicated mathematical objects by making <em>conjectures</em> (ideas about how those patterns might work), which are promoted to <em>theorems</em> if we can prove they are true. This process relies on our intuition as much as our knowledge. </p>
<p>Over the past few years I’ve been working with experts at artificial intelligence (AI) company DeepMind to find out whether their programs can help with the creative or intuitive aspects of mathematical research. In a new paper <a href="https://www.nature.com/articles/s41586-021-04086-x">published in Nature</a>, we show they can: recent techniques in AI have been essential to the discovery of a new conjecture and a new theorem in two fields called “knot theory” and “representation theory”.</p>
<h2>Machine intuition</h2>
<p>Where does the intuition of a mathematician come from? One can ask the same question in any field of human endeavour. How does a chess grandmaster know their opponent is in trouble? How does a surfer know where to wait for a wave? </p>
<p>The short answer is we don’t know. Something miraculous seems to happen in the human brain. Moreover, this “miraculous something” takes thousands of hours to develop and is not easily taught.</p>
<p>The past decade has seen computers display the first hints of something like human intuition. The most striking example of this occurred in 2016, in a Go match between DeepMind’s AlphaGo program and Lee Sedol, one of the world’s best players.</p>
<p>AlphaGo won 4–1, and experts observed that some of AlphaGo’s moves displayed human-level intuition. One particular move (<a href="https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-redefined-future/">“move 37”</a>) is now famous as a new discovery in the game. </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/HT-UZkiOLv8?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
</figure>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/ai-has-beaten-us-at-go-so-what-next-for-humanity-55945">AI has beaten us at Go. So what next for humanity?</a>
</strong>
</em>
</p>
<hr>
<h2>How do computers learn?</h2>
<p>Behind these breakthroughs lies a technique called deep learning. On a computer one builds a neural network – essentially a crude mathematical model of a brain, with many interconnected neurons. </p>
<p>At first, the network’s output is useless. But over time (from hours to even weeks or months), the network is trained, essentially by adjusting the firing rates of the neurons.</p>
<p>Such ideas were tried in the 1970s with unconvincing results. Around 2010, however, <a href="https://www.nature.com/articles/nature14539">a revolution occurred</a> when researchers drastically increased the number of neurons in the model (from hundreds in the 1970s to billions today).</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=745&fit=crop&dpr=1 600w, https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=745&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=745&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=936&fit=crop&dpr=1 754w, https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=936&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/434870/original/file-20211130-17-1ydvy5w.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=936&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">One of the first neural networks, the Mark I Perceptron, was built in the 1950s. The goal was to classify digital images, but results were disappointing.</span>
<span class="attribution"><span class="source">Cornell University</span></span>
</figcaption>
</figure>
<p>Traditional computer programs struggle with many tasks humans find easy, such as natural language processing (reading and interpreting text), and speech and image recognition. </p>
<p>With the deep learning revolution of the 2010s, computers began performing well on these tasks. AI has essentially brought vision and speech to machines.</p>
<p>Training neural nets requires huge amounts of data. What’s more, trained deep learning models often function as “black boxes”. We know they often give the right answer, but we usually don’t know (and can’t ascertain) why.</p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=450&fit=crop&dpr=1 600w, https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=450&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=450&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=566&fit=crop&dpr=1 754w, https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=566&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/434865/original/file-20211130-17-1d5nahs.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=566&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Deep learning systems often function as ‘black boxes’: data goes in and data comes out, but we have difficulty making sense of what happens in between.</span>
<span class="attribution"><span class="source">Shutterstock</span></span>
</figcaption>
</figure>
<h2>A lucky encounter</h2>
<p>My involvement with AI began in 2018, when I was elected a Fellow of the Royal Society. At the induction ceremony in London I met Demis Hassabis, chief executive of DeepMind. </p>
<p>Over a coffee break we discussed deep learning, and possible applications in mathematics. Could machine learning lead to discoveries in mathematics, like it had in Go?</p>
<p>This fortuitous conversation led to my collaboration with the team at DeepMind. </p>
<p>Mathematicians like myself often use computers to check or perform long computations. However, computers usually cannot help me develop intuition or suggest a possible line of attack. So we asked ourselves: can deep learning help mathematicians build intuition?</p>
<p>With the team from DeepMind, we trained models to predict certain quantities called Kazhdan-Lusztig polynomials, which I have spent most of my mathematical life studying. </p>
<p>In my field we study <em>representations</em>, which you can think of as being like molecules in chemistry. In much the same way that molecules are made of atoms, the make up of representations is governed by Kazhdan-Lusztig polynomials.</p>
<p>Amazingly, the computer was able to predict these Kazhdan-Lusztig polynomials with incredible accuracy. The model seemed to be onto something, but we couldn’t tell what. </p>
<p>However, by “peeking under the hood” of the model, we were able to find a clue which led us to a new conjecture: that Kazhdan-Lusztig polynomials can be distilled from a much simpler object (a mathematical graph).</p>
<p>This conjecture suggests a way forward on a problem that has stumped mathematicians for more than 40 years. Remarkably, for me, the model was providing intuition!</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/how-explainable-artificial-intelligence-can-help-humans-innovate-151737">How explainable artificial intelligence can help humans innovate</a>
</strong>
</em>
</p>
<hr>
<p>In parallel work with DeepMind, mathematicians Andras Juhasz and Marc Lackenby at the University of Oxford used similar techniques to discover a new theorem in the mathematical field of knot theory. The theorem gives a relation between traits (or “invariants”) of knots that arise from different areas of the mathematical universe.</p>
<p>Our paper reminds us that intelligence is not a single variable, like the result of an IQ test. Intelligence is best thought of as having many dimensions. </p>
<p>My hope is that AI can provide another dimension, deepening our understanding of the mathematical world, as well as the world in which we live.</p><img src="https://counter.theconversation.com/content/172900/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Geordie Williamson is a Professor at the University of Sydney, and a consultant in Pure Mathematics for DeepMind, a subsidiary of Alphabet.</span></em></p>Machine learning systems can now aid human intuition to create mathematical conjectures and proofs.Geordie Williamson, Professor of Mathematics, University of SydneyLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1511812020-12-02T13:28:54Z2020-12-02T13:28:54ZAI makes huge progress predicting how proteins fold – one of biology’s greatest challenges – promising rapid drug development<figure><img src="https://images.theconversation.com/files/372322/original/file-20201201-15-s2hltf.png?ixlib=rb-1.1.0&rect=5%2C2%2C973%2C431&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">A simple chain of amino acids folds into a complex three-dimensional structure.</span> </figcaption></figure><p><strong>Takeaways</strong></p>
<ul>
<li><p><strong>A “deep learning” software program from Google-owned lab DeepMind showed great progress in solving one of biology’s greatest challenges – understanding protein folding.</strong> </p></li>
<li><p><strong>Protein folding is the process by which a protein takes its shape from a string of building blocks to its final three-dimensional structure, which determines its function.</strong></p></li>
<li><p><strong>By better predicting how proteins take their structure, or “fold,” scientists can more quickly develop drugs that, for example, block the action of crucial viral proteins.</strong> </p></li>
</ul>
<hr>
<p>Solving what biologists call “the protein-folding problem” is a big deal. Proteins are the workhorses of cells and are present in all living organisms. They are made up of long chains of amino acids and are vital for the structure of cells and communication between them as well as regulating all of the chemistry in the body. </p>
<p>This week, the Google-owned artificial intelligence company <a href="https://www.deepmind.com">DeepMind</a> demonstrated a deep-learning program called <a href="https://deepmind.com/blog/article/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology">AlphaFold2</a>, which experts are calling a <a href="https://www.nature.com/articles/d41586-020-03348-4">breakthrough</a> toward solving the grand challenge of <a href="https://doi.org/10.1038/d41586-020-03348-4">protein folding</a>. </p>
<p>Proteins are long chains of amino acids linked together like beads on a string. But for a protein to do its job in the cell, it must “fold” – a process of twisting and bending that transforms the molecule into a complex three-dimensional structure that can interact with its target in the cell. If the folding is disrupted, then the protein won’t form the correct shape – and it won’t be able to perform its job inside the body. This can lead to disease – as is the case in a common disease like Alzheimer’s, and rare ones like cystic fibrosis.</p>
<p>Deep learning is a computational technique that uses the often hidden information contained in vast datasets to solve questions of interest. It’s been used widely in fields such as games, speech and voice recognition, autonomous cars, science and medicine.</p>
<p>I believe that tools like AlphaFold2 will help scientists to design new types of proteins, ones that may, for example, help break down plastics and fight future viral pandemics and disease. </p>
<p><a href="https://scholar.google.com/citations?user=RpiSPiwAAAAJ&hl=en">I am a computational chemist</a> and author of the book <a href="https://rowman.com/ISBN/9781633886407/The-State-of-Science-What-the-Future-Holds-and-the-Scientists-Making-It-Happen">The State of Science</a>. My students and I study the structure and properties of <a href="https://www.conncoll.edu/ccacad/zimmer/GFP-ww/GFP-1.htm">fluorescent proteins</a> using protein-folding computer programs based on classical physics. </p>
<p>After decades of study by thousands of research groups, these protein-folding prediction programs are very good at calculating structural changes that occur when we make small alterations to known molecules. </p>
<p>But they haven’t adequately managed to predict how proteins fold from scratch. Before deep learning came along, the protein-folding problem seemed impossibly hard, and it seemed poised to frustrate computational chemists for many decades to come.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=510&fit=crop&dpr=1 600w, https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=510&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=510&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=641&fit=crop&dpr=1 754w, https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=641&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/372313/original/file-20201201-23-12msmry.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=641&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">A chain of amino acids goes through several folding steps, which occurs through hydrogen bonds between amino acids in different regions of the protein, before arriving at the final structure. The example shown here is hemoglobin, a protein in red blood cells that transports oxygen to body tissues.</span>
<span class="attribution"><a class="source" href="https://upload.wikimedia.org/wikipedia/commons/2/26/225_Peptide_Bond-01.jpg">Anatomy & Physiology, Connexions website</a>, <a class="license" href="http://creativecommons.org/licenses/by/4.0/">CC BY</a></span>
</figcaption>
</figure>
<h2>Protein folding</h2>
<p>The sequence of the amino acids – which is encoded in DNA – defines the protein’s 3D shape. The shape determines its function. If the structure of the protein changes, it is unable to perform its function. Correctly predicting protein folds based on the amino acid sequence could revolutionize drug design, and explain the causes of new and old diseases. </p>
<p>All proteins with the same sequence of amino acid building blocks fold into the same three-dimensional form, which optimizes the interactions between the amino acids. They do this within milliseconds, although they have an astronomical number of possible configurations available to them – <a href="https://web.archive.org/web/20110523080407/http://www-miller.ch.cam.ac.uk/levinthal/levinthal.html">about 10 to the power of 300</a>. This massive number is what makes it hard to predict how a protein folds even when scientists know the full sequence of amino acids that go into making it. Previously predicting the structure of protein from the amino acid sequence was impossible. Protein structures were experimentally determined, a time-consuming and expensive endeavor. </p>
<p>Once researchers can better predict how proteins fold, they’ll be able to better understand how cells function and how misfolded proteins cause disease. Better protein prediction tools will also help us design drugs that can target a particular topological region of a protein where chemical reactions take place. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=300&fit=crop&dpr=1 600w, https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=300&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=300&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=377&fit=crop&dpr=1 754w, https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=377&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/372314/original/file-20201201-23-86jeuv.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=377&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">What’s your move?</span>
<span class="attribution"><a class="source" href="https://www.gettyimages.com/detail/photo/robot-hand-chessboard-royalty-free-image/1255171787?adppopup=true">style-photography/Getty Images</a></span>
</figcaption>
</figure>
<h2>AlphaFold is born from deep-learning chess, Go and poker games</h2>
<p>The success of DeepMind’s protein-folding prediction program, called <a href="https://deepmind.com/research/case-studies/alphafold">AlphaFold</a>, is not unexpected. Other deep-learning programs written by <a href="https://deepmind.com/about">DeepMind</a> have demolished the world’s best chess, Go and poker players.</p>
<p>In 2016 <a href="https://www.chessprogramming.org/Stockfish">Stockfish-8</a>, an open-source chess engine, was the world’s computer chess champion. It evaluated 70 million chess positions per second and had centuries of accumulated human chess strategies and decades of computer experience to draw upon. It played efficiently and brutally, mercilessly beating all its human challengers without an ounce of finesse. Enter deep learning. </p>
<p>On Dec. 7, 2017, Google’s deep-learning chess program <a href="http://doi.org/10.1126/science.aar6404">AlphaZero</a> thrashed Stockfish-8. The chess engines played 100 games, with AlphaZero winning 28 and tying 72. It didn’t lose a single game. AlphaZero did only 80,000 calculations per second, as opposed to Stockfish-8’s 70 million calculations, and it took just four hours to learn chess from scratch by playing against itself a few million times and optimizing its neural networks as it learned from its experience. </p>
<p><a href="https://web.stanford.edu/%7Esurag/posts/alphazero.html">AlphaZero</a> didn’t learn anything from humans or chess games played by humans. It taught itself and, in the process, derived strategies never seen before. In a <a href="https://doi.org/10.1126/science.aaw2221">commentary</a> in Science magazine, former world chess champion Garry Kasparov wrote that by learning from playing itself, AlphaZero developed strategies that “reflect the truth” of chess rather than reflecting “the priorities and prejudices” of the programmers. “It’s the embodiment of the cliché ‘work smarter, not harder.’” </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/gg7WjuFs8F4?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">How do proteins fold?</span></figcaption>
</figure>
<h2>CASP – the Olympics for molecular modelers</h2>
<p>Every two years, the world’s top computational chemists test the abilities of their programs to predict the folding of proteins and compete in the <a href="https://predictioncenter.org">Critical Assessment of Structure Prediction</a> (CASP) competition. </p>
<p>In the competition, teams are given the linear sequence of amino acids for about 100 proteins for which the 3D shape is known but hasn’t yet been published; they then have to compute how these sequences would fold. In 2018 AlphaFold, the deep-learning rookie at the competition, beat all the traditional programs – but barely. </p>
<p>Two years later, on Monday, it was announced that Alphafold2 had won the 2020 competition by a healthy margin. It whipped its competitors, and its predictions were comparable to the existing experimental results determined through gold standard techniques like X-ray diffraction crystallography and cryo-electron microscopy. Soon I expect AlphaFold2 and its progeny will be the methods of choice to determine protein structures before resorting to experimental techniques that require painstaking, laborious work on expensive instrumentation.</p>
<p>One of the reasons for AlphaFold2’s success is that it could use the <a href="https://www.rcsb.org/">Protein Database</a>, which has over 170,000 experimentally determined 3D structures, to train itself to calculate the correctly folded structures of proteins. </p>
<p>The potential impact of AlphaFold can be appreciated if one compares the number of all published protein structures – approximately 170,000 – with the 180 million DNA and protein sequences deposited in the <a href="https://www.uniprot.org">Universal Protein Database</a>. AlphaFold will help us sort through treasure troves of DNA sequences hunting for new proteins with unique structures and <a href="https://deepmind.com/blog/article/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology">functions</a>.</p>
<h2>Has AlphaFold made me, a molecular modeler, redundant?</h2>
<p>As with the chess and Go programs – AlphaZero and AlphaGo – we don’t exactly know what the AlphaFold2 algorithm is doing and why it uses certain correlations, but we do know that it works. </p>
<p>Besides helping us predict the structures of important proteins, understanding AlphaFold’s “thinking” will also help us gain new insights into the mechanism of protein folding.</p>
<p>[<em>Deep knowledge, daily.</em> <a href="https://theconversation.com/us/newsletters/the-daily-3?utm_source=TCUS&utm_medium=inline-link&utm_campaign=newsletter-text&utm_content=deepknowledge">Sign up for The Conversation’s newsletter</a>.]</p>
<p>One of the most common fears expressed about AI is that it will lead to large-scale unemployment. AlphaFold still has a significant way to go before it can consistently and successfully predict protein folding. </p>
<p>However, once it has matured and the program can simulate protein folding, computational chemists will be integrally involved in improving the programs, trying to understand the underlying correlations used, and applying the program to solve important problems such as the protein misfolding associated with many diseases such as Alzheimer’s, Parkinson’s, cystic fibrosis and Huntington’s disease. </p>
<p>AlphaFold and its offspring will certainly change the way computational chemists work, but it won’t make them redundant. Other areas won’t be as fortunate. In the past robots were able to replace humans doing manual labor; with AI, our cognitive skills are also being challenged.</p><img src="https://counter.theconversation.com/content/151181/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Marc Zimmer does not work for, consult, own shares in or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>Scientists in an artificial intelligence lab have made a breakthrough in solving the problem of how proteins fold into their final three-dimensional shape. The work could speed up creation of drugs.Marc Zimmer, Professor of Chemistry, Connecticut CollegeLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1261042019-10-31T04:12:20Z2019-10-31T04:12:20ZRobots can outwit us on the virtual battlefield, so let’s not put them in charge of the real thing<figure><img src="https://images.theconversation.com/files/299622/original/file-20191031-187934-1axfejq.png?ixlib=rb-1.1.0&rect=15%2C5%2C3339%2C2082&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">DeepMind's artificial intelligence-powered AlphaStar (green) repels an attack in the virtual world of StarCraft II.</span> <span class="attribution"><span class="source">DeepMind</span></span></figcaption></figure><p>Artificial intelligence developer DeepMind has just <a href="https://www.nature.com/articles/s41586-019-1724-z">announced</a> its latest milestone: a bot called AlphaStar that plays the popular real-time strategy game StarCraft II at Grandmaster level. </p>
<p>This isn’t the first time a bot has outplayed humans in a strategy war game. In 1981, a program called <a href="https://www.sciencedirect.com/science/article/abs/pii/S0004370283800058">Eurisko</a>, developed by artificial intelligence (AI) pioneer Doug Lenat, won the US championship of Traveller, a highly complex strategy war game in which players design a fleet of 100 ships. Eurisko was consequently made an honorary Admiral in the Traveller navy. </p>
<p>The following year, the tournament rules were overhauled in an attempt to thwart computers. But Eurisko triumphed for a second successive year. With officials threatening to abolish the tournament if a computer won again, Lenat retired his program.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/if-machines-can-beat-us-at-games-does-it-make-them-more-intelligent-than-us-60555">If machines can beat us at games, does it make them more intelligent than us?</a>
</strong>
</em>
</p>
<hr>
<p>DeepMind’s PR department would have you believe that StarCraft “has emerged by consensus as the next grand challenge (in computer games)” and “has been a grand challenge for AI researchers for over 15 years”. </p>
<p>In the most recent StarCraft computer game tournament, only four entries came from academic or industrial research labs. The nine other bots involved were written by lone individuals outside the mainstream of AI research. </p>
<p>In fact, the 42 authors of DeepMind’s paper, <a href="https://www.nature.com/articles/s41586-019-1724-z">published today in Nature</a>, greatly outnumber the rest of the world building bots for StarCraft. Without wishing to take anything away from an impressive feat of collaborative engineering, if you throw enough resources at a problem, success is all but assured.</p>
<p>Unlike recent successes with computer chess and <a href="https://theconversation.com/googles-new-go-playing-ai-learns-fast-and-even-thrashed-its-former-self-85979">Go</a>, AlphaStar didn’t learn to outwit humans simply by playing against itself. Rather, it learned by imitating the best bits from nearly a million games played by top-ranked human players. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/googles-new-go-playing-ai-learns-fast-and-even-thrashed-its-former-self-85979">Google’s new Go-playing AI learns fast, and even thrashed its former self</a>
</strong>
</em>
</p>
<hr>
<p>Without this input, AlphaStar was beaten convincingly by 19 out of 20 human players on the StarCraft game server. AlphaStar also played anonymously on that server so that humans couldn’t exploit any weaknesses that might have been uncovered in earlier games. </p>
<p>AlphaStar did beat Grzegorz “MaNa” Komincz, one of the world’s top professional StarCraft players, <a href="https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii">in December last year</a>. But this was a version of AlphaStar with much faster reflexes than any human, and unlimited vision of the playing board (unlike human players who can only see a portion of it at any one time). This was hardly a level playing field.</p>
<p>Nevertheless, StarCraft does have some features that makes AlphaStar an impressive advance, if not truly a breakthrough. Unlike chess or Go, players in StarCraft have imperfect information about the state of play, and the set of possible actions you can make at any point is much larger. And StarCraft unfolds in real time and requires long-term planning.</p>
<h2>Robot wars</h2>
<p>This raises the question of whether, in the future, we will see robots not just fighting wars but planning them too. Actually, we already have both. </p>
<p>Despite the many warnings raised by AI researchers such as myself – as well as by founders of AI and robotics companies, Nobel Peace Laureates, and church leaders – fully autonomous weapons, also known as “killer robots”, have been developed and will soon be used. </p>
<p>In 2020, Turkey will <a href="http://ehamedya.com/turkey-to-start-operating-kamikaze-drones-near-syrian-border-in-2020_28626.html">deploy kamikaze drones</a> on its border with Syria. These drones will use computer vision to identify, track and kill people without human intervention.</p>
<p>This is a terrible development. Computers do not have the moral capability to decide who lives or dies. They have neither empathy nor compassion. “Killer robots” will change the very nature of conflict for the worse.</p>
<p>As for “robot generals”, computers have been helping generals plan war for decades.</p>
<p>In Desert Storm, during the Gulf War of the early 1990s, AI scheduling tools were used to plan the buildup of forces in the Middle East prior to conflict. A US general told me shortly afterwards that the amount of money saved by doing this was equivalent to everything that had been spent on AI research until then. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=382&fit=crop&dpr=1 600w, https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=382&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=382&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=480&fit=crop&dpr=1 754w, https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=480&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/299623/original/file-20191031-187934-fadi53.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=480&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">US fighters flying over Kuwait in 1991. Positioning military hardware is complex and costly.</span>
<span class="attribution"><span class="source">US Air Force</span></span>
</figcaption>
</figure>
<p>Computers have also been used extensively by generals to war-game potential strategies. But just as we wouldn’t entrust all battlefield decisions to a single soldier, handing over the full responsibilities of a general to a computer would be a step too far. </p>
<p>Machines cannot be held accountable for their decisions. Only humans can be. This is a cornerstone of international humanitarian law. </p>
<p>Nevertheless, to cut through the fog of war and deal with the vast amount of information flowing back from the front, generals will increasingly rely on computer support in their decision-making. </p>
<p>If this results in fewer civilian deaths, less friendly fire, and more respect for international humanitarian law, we should welcome such computer assistance. But the buck needs to stop with humans, not machines.</p>
<p>Here’s a final question to ponder. If tech companies like Google really don’t want us to worry about computers taking over, why are they building bots to win virtual wars rather than concentrating on, say, more peaceful e-sports? With all due respect to sports fans, the stakes would be much lower.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/robots-will-be-fifa-champions-if-they-keep-their-eyes-on-the-ball-11238">Robots will be FIFA champions – if they keep their eyes on the ball</a>
</strong>
</em>
</p>
<hr>
<img src="https://counter.theconversation.com/content/126104/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Toby Walsh does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>StarCraft II is the latest complex game to be conquered by artificial intelligence. But if robots now reign supreme at virtual war, where does that leave us when it comes to real conflict?Toby Walsh, Professor of AI at UNSW, Research Group Leader, Data61Licensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/1212012019-07-31T11:55:44Z2019-07-31T11:55:44ZTo create a sustainable NHS we must urgently invest in cutting-edge technology<figure><img src="https://images.theconversation.com/files/286233/original/file-20190730-186833-g6r4zq.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Health secretary Matt Hancock understands the importance of technology in healthcare.</span> <span class="attribution"><a class="source" href="http://www.epa.eu/politics-photos/government-photos/british-prime-minister-boris-johnson-s-government-members-photos-55365270">Neil Hall/EPA</a></span></figcaption></figure><p>While Theresa May was counting down the hours to the end of her premiership, the government unexpectedly slipped out a <a href="https://www.gov.uk/government/consultations/advancing-our-health-prevention-in-the-2020s">green paper</a> on preventing ill health in a last-minute shelf-clearing operation.</p>
<p>Much of the comment it attracted has focused on proposed new curbs on the <a href="https://www.bmj.com/content/366/bmj.l4829">food</a> and <a href="https://www.bbc.co.uk/news/health-49079515">tobacco</a> industries, after Boris Johnson had earlier signalled his opposition to “sin taxes” during the Tory leadership contest. In fact, the incoming prime minister supported such taxes as London mayor when <a href="https://www.healthylondon.org/resource/better-health-london-report/">I proposed them in 2014</a> as chair of the London Health Commission.</p>
<p>But a key part of the green paper has so far gone unreported. Ordered by the digitally savvy health secretary, Matt Hancock, it advocates a big increase in the role of technology in preventing ill health. This will be critical for creating a sustainable NHS. </p>
<p>We have <a href="https://www.cam.ac.uk/research/news/ounce-of-prevention-pound-of-cure">known for centuries</a> that “an ounce of prevention is worth a pound of cure”, yet blanket exhortations such as “eat less and exercise more” have had little impact on the nation’s health. Prevention and public health have remained neglected parts of the NHS, taking less than 5% of the overall budget and suffering a fall in spending of £600m since 2014-15, <a href="https://www.ippr.org/research/publications/ending-the-blame-game">according to the Institute of Public Policy Research</a>. </p>
<p>That must change. With the latest developments in genomics, artificial intelligence (AI) and digital innovation, we will increasingly be able to intervene earlier in the disease process with personalised plans, tailored to individuals, delivering targeted support and specific protection against future threats.</p>
<p>We will be able to use the huge quantities of data amassed by the NHS to predict who will be ill, with which disease, and what interventions will be most effective – with safeguards to guarantee patient confidentiality.</p>
<h2>Intelligent investments</h2>
<p>But this will require ministers to commit to bold new investment – an investment that is vital not only for the future of the NHS but also for the future of life sciences in the UK. </p>
<p>Britain has led the way in the development of genomics – understanding how our genetic inheritance affects both our risk of disease and our response to treatment – with initiatives such as the <a href="https://www.genomicsengland.co.uk/about-genomics-england/the-100000-genomes-project/">100,000 Genomes Project</a>. The latest advances are making a real difference, with up to <a href="https://www.gov.uk/government/consultations/advancing-our-health-prevention-in-the-2020s/advancing-our-health-prevention-in-the-2020s-consultation-document">half of cancer patients</a> identifiable as suitable for a different treatment pathway and <a href="https://www.gov.uk/government/consultations/advancing-our-health-prevention-in-the-2020s/advancing-our-health-prevention-in-the-2020s-consultation-document">one in four patients</a> with rare diseases who were previously without a diagnosis now receiving one.</p>
<p>Developments in AI and machine learning also promise big gains in areas such as health screening. For example, a collaboration between Moorfields Eye Hospital in London and tech company DeepMind has led to the creation of an AI system that is able to diagnose <a href="https://www.gov.uk/government/consultations/advancing-our-health-prevention-in-the-2020s/advancing-our-health-prevention-in-the-2020s-consultation-document">more than 50 different eye diseases with 94% accuracy</a>, matching the best human experts and increasing the potential scope of screening programmes. </p>
<p>Digital versions of the diabetes prevention programme, providing personal health coaching and dietary recommendations, are already reaching more people through <a href="https://www.gov.uk/government/consultations/advancing-our-health-prevention-in-the-2020s/advancing-our-health-prevention-in-the-2020s-consultation-document">wearable technologies, apps and websites</a>, helping to prevent the 3m people at high risk of diabetes joining the 10m already dependent on drugs and associated NHS treatment. Digital technologies are being used to support other patients, helping them access the resources they need to manage their own conditions.</p>
<p>These are advances, in some cases already being realised, that must be extended and spread across the NHS. But to attract the necessary investment we need to reframe the healthcare challenge.</p>
<p>The NHS continues to treat people when they fall ill. That is far too late. It is also unsustainable. With an ageing population and a growing burden of chronic disease, we are spending billions of pounds on avoidable illness.</p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/286236/original/file-20190730-186829-e1f1pi.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">An ounce of prevention.</span>
<span class="attribution"><a class="source" href="https://www.shutterstock.com/download/confirm/741180823?src=tBGblTPwzoTSI1ECnQiCfA-1-85&studio=1&size=medium_jpg">LanaG/Shutterstock</a></span>
</figcaption>
</figure>
<h2>Don’t wait till the enemy is at the door</h2>
<p>Disease poses a threat to our health security in the same way that an invading army is a threat to our national security, as the US biotech entrepreneur <a href="https://youtu.be/uv-18VffraU">Noubar Afeynan has suggested</a>. </p>
<p>We do not wait for enemy forces to gather on our shores before responding with military intervention. We rely on our intelligence and diplomatic services to alert us when threats exist and keep them neutralised.</p>
<p>Similarly, drug treatment and surgery should be seen as the last line of defence, not the first. Our immune system protects against ill health and, as it deteriorates over time, we encounter a series of health challenges to our security – cardiovascular, neurological, oncological, infectious. </p>
<p>Prevention and detection then become key to preserving our health security, shifting the focus of activity from last-line treatment to continuous monitoring and intervention, providing ongoing care. </p>
<p>Reframing the health challenge in this way could provide the necessary impetus for change. We must seize this opportunity to transform medicine, by harnessing the power of technology to deliver intelligent prevention, enabling us to spot disease earlier, intervene sooner and save lives. </p>
<p>This shift is essential to help protect us from future disease threats. But it will also be vital to help sustain the NHS to continue providing care for future generations.</p><img src="https://counter.theconversation.com/content/121201/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Lord Darzi holds the Paul Hamlyn Chair of Surgery at Imperial College London and the director of the institute of Global Health Innovation. Lord Darzi Is the principal investigator of larger number of grant from the Wellocme Trust, EPSRC, Cancer research UK as few examples. He is a board member of NHS Improvement. Lord Darzi also chairs the Accelerated Access Collaborative reporting to the Secretary of State for health. Lord Darzi also member of Evelo Board; a biotech company focusing on the role of Microbial on gut -body network and also chairs the scientific advisory board of Deepmind. Lord Darzi led a major review of NHS reform as the Parliamentary Under secretary of Health and published the seminal report High Quality Care for all in 2008. In 2014 he chaired the London Health Commission for then the Mayor of London Mr Boris Johnson.</span></em></p>To be sustainable, the NHS needs to invest in AI and other advanced technologies.Ara Darzi, Director of the Institute of Global Health Innovation, Imperial College LondonLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/944372018-04-05T14:44:14Z2018-04-05T14:44:14ZTo drive AI forward, teach computers to play old-school text adventure games<figure><img src="https://images.theconversation.com/files/214305/original/file-20180411-543-1dcho2z.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Ready player one?</span> <span class="attribution"><a class="source" href="https://commons.wikimedia.org/wiki/File:Colossal_Cave_Adventure_on_VT100_terminal.jpg#/media/File:Colossal_Cave_Adventure_on_VT100_terminal.jpg">Wikimedia</a></span></figcaption></figure><p>Games have long been used as test beds and benchmarks for artificial intelligence, and there has been no shortage of achievements in recent months. Google DeepMind’s <a href="https://theconversation.com/googles-latest-go-victory-shows-machines-are-no-longer-just-learning-theyre-teaching-78410">AlphaGo</a> and <a href="https://www.theregister.co.uk/2017/12/19/poker_bot_libratus_ai/">poker bot Libratus</a> from Carnegie Mellon University have both beaten human experts at games that have traditionally been hard for AI – some 20 years after IBM’s DeepBlue achieved the same feat <a href="https://www.theguardian.com/theguardian/2011/may/12/deep-blue-beats-kasparov-1997">in chess</a>. </p>
<p>Games like these have the attraction of clearly defined rules; they are relatively simple and cheap for AI researchers to work with, and they provide a variety of cognitive challenges at any desired level of difficulty. By inventing algorithms that play them well, researchers hope to gain insights into the mechanisms needed to function autonomously. </p>
<p>With the arrival of the latest techniques in AI and machine learning, attention is <a href="https://project.dke.maastrichtuniversity.nl/cig2018/?page_id=255">now shifting</a> to visually detailed computer games – including the 3D shooter Doom, <a href="https://github.com/mgbellemare/Arcade-Learning-Environment">various 2D Atari games</a> such as Pong and Space Invaders, and the real-time strategy game StarCraft. </p>
<p>This is all certainly progress, but a key part of the bigger AI picture is being overlooked. Research has prioritised games in which all the actions that can be performed are known in advance, be it moving a knight or firing a weapon. The computer is given all the options from the outset and the focus is on how well it chooses between them. The problem is that this disconnects AI research from the task of making computers genuinely autonomous. </p>
<h2>Banana skins</h2>
<p>Getting computers to determine which actions even exist in a given context presents conceptual and practical challenges which games researchers have barely attempted to resolve so far. The “monkey and bananas” problem is one example of a longstanding AI conundrum in which no recent progress has been made. </p>
<figure class="align-right zoomable">
<a href="https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=237&fit=clip" srcset="https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=690&fit=crop&dpr=1 600w, https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=690&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=690&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=867&fit=crop&dpr=1 754w, https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=867&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/213170/original/file-20180404-189807-zzpsqv.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=867&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Headscratcher.</span>
<span class="attribution"><a class="source" href="https://www.shutterstock.com/image-photo/monkey-man-holding-banana-over-colorful-283183991?src=UaQrgHzv4Gm6OcQmcC2fzA-1-42">Luis Molinero</a></span>
</figcaption>
</figure>
<p>The problem was <a href="https://www.sciencedirect.com/science/article/pii/S0004370210001827">originally posed</a> by John McCarthy, one of the founding fathers of AI, in 1963: there is a room containing a chair, a stick, a monkey and a bunch of bananas hanging on a ceiling hook. The task is for a computer to come up with a sequence of actions to enable the monkey to acquire the bananas. </p>
<p>McCarthy made a key distinction between two aspects of this task in terms of artificial intelligence. Physical feasibility – determining whether a particular sequence of actions is physically realisable; and epistemic or knowledge-related feasibility – determining which possible actions for the monkey actually exist. </p>
<p>Determining what is physically feasible for the monkey is very easy for a computer if it is told all the possible actions in advance – “climb on chair”, “wave stick” and so forth. A simple program that instructs the computer to go through all the possible sequences of actions one by one will quickly arrive at the best solution. </p>
<p>If the computer has to first determine which actions are even possible, however, it is a much tougher challenge. It raises questions about how we represent knowledge, the necessary and sufficient conditions of knowing something, and how we know when enough knowledge has been acquired. In highlighting these problems, McCarthy <a href="https://dl.acm.org/citation.cfm?id=216000">said</a>:</p>
<blockquote>
<p>Our ultimate objective is to make programs that learn from their experience as effectively as humans do.</p>
</blockquote>
<p>Until computers can tackle problems without any predetermined description of possible actions, this objective can’t be achieved. It is unfortunate that AI researchers are neglecting this: not only are these problems harder and more interesting, they look like a prerequisite for making further meaningful progress in the field. </p>
<h2>Text appeal</h2>
<p>To operate autonomously in a complex environment, it is impossible to describe in advance how best to manipulate – or even characterise – the objects there. Teaching computers to get around these difficulties immediately leads to deep questions about learning from previous experience.</p>
<p>Rather than focusing on games like Doom or StarCraft, where it is possible to avoid this problem, a more promising test for modern AI could be the humble text adventure from the 1970s and 1980s. </p>
<p>In the days before computers had sophisticated graphics capabilities, games like Colossal Cave and Zork were popular. Players were told about their environment by messages on the screen:</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=376&fit=crop&dpr=1 600w, https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=376&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=376&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=473&fit=crop&dpr=1 754w, https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=473&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/213173/original/file-20180404-189816-e49be.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=473&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">Picture this.</span>
</figcaption>
</figure>
<p>They had to respond with simple instructions, usually in the form of a verb or a verb plus a noun – “look”, “take box” and so on. Part of the challenge was to work out which actions were possible and useful and to respond accordingly. </p>
<p>A good challenge for modern AI would be to take on the role of a player in such an adventure. The computer would have to make sense of the text descriptions on the screen and respond to them with actions, using some predictive mechanism to determine their likely effect. </p>
<p>More sophisticated behaviours on part of the computer would involve exploring the environment, defining goals, making goal-oriented action choices and solving the various intellectual challenges typically required to progress. </p>
<p>How well modern AI methods of the kind promoted by tech giants like IBM, Google, Facebook or Microsoft would fare in these text adventures is an open question – as is how much specialist human knowledge they would require for each new scenario. </p>
<p>To measure progress in this area, for the past two years we <a href="http://atkrye.github.io/IEEE-CIG-Text-Adventurer-Competition/2018/01/16/announceThirdYear/">have been running a competition</a> at the IEEE Conference on Computational Intelligence and Games, which <a href="https://project.dke.maastrichtuniversity.nl/cig2018/">this year takes place in Maastricht</a> in the Netherlands in August. Competitors submit entries in advance, and can use the AI technology of their choice to build programs that can play these games by making sense of a text description and outputting appropriate text commands in return. </p>
<p>In short, researchers need to reconsider their priorities if AI is to keep progressing. If unearthing the discipline’s neglected roots turns out to be fruitful, the monkey may finally gets his bananas after all.</p><img src="https://counter.theconversation.com/content/94437/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.</span></em></p>It’s time programmers looked out old computer text adventures like Zork and Colossal Cave from the 1970s and 1980s.Jerry Swan, Senior Research Fellow, University of YorkHendrik Baier, Research Associate for Artificial Intelligence and Data Analytics, University of YorkTimothy Atkinson, Doctoral Researcher, University of YorkLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/887072017-12-06T15:49:33Z2017-12-06T15:49:33ZDeepMind: can we ever trust a machine to diagnose cancer?<figure><img src="https://images.theconversation.com/files/197950/original/file-20171206-917-yn6wxu.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption"></span> <span class="attribution"><a class="source" href="https://www.shutterstock.com/download/confirm/599551640?src=BrNSLwC50QUvNbSJWD6CGQ-1-33&size=huge_jpg">Shutterstock</a></span></figcaption></figure><p>DeepMind has recently announced a fresh collaborative partnership with the UK’s health service, with plans for the artificial intelligence firm to develop machine learning technology to research breast cancer.</p>
<p>DeepMind, a Google subsidiary, is perhaps best known for successfully building AI that is now <a href="https://storage.googleapis.com/deepmind-media/alphago/AlphaGoNaturePaper.pdf">better than humans at the ancient game of Go</a>. But in recent months – when attempting to apply this tech to serious healthcare issues – it has been on the sidelines of a data breach storm.</p>
<p>In July, DeepMind’s collaboration with London’s Royal Free hospital led to the NHS trust violating the UK’s data protection laws. </p>
<p>The <a href="https://ico.org.uk/about-the-ico/news-and-events/news-and-blogs/2017/07/royal-free-google-deepmind-trial-failed-to-comply-with-data-protection-law/">Information Commissioner’s Office (ICO)</a> found that Royal Free’s decision to share 1.6m personally identifiable patient records with DeepMind for the development of Streams – an automated kidney injury detection software – was “legally inappropriate”. DeepMind wasn’t directly criticised by the ICO.</p>
<p>Personal records included patients’ HIV-positive status, as well as details of drug overdoses and abortions. Royal Free’s breach generated considerable media attention at the time, and it means that DeepMind’s latest partnership with an NHS trust will be scrutinised carefully.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/F1wlCerC40E?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
</figure>
<p>It will be working with Cancer Research UK, Imperial College London and the Royal Surrey NHS trust to apply <a href="https://deepmind.com/blog/applying-machine-learning-mammography/">machine learning to mammography screening for breast cancer</a>. This is a laudable aim, and one to be taken very seriously, given DeepMind’s track record. London-based DeepMind emerged from academic research, assisted by Google’s deep pockets. It is now owned by Google’s parent company Alphabet. </p>
<p>Its success has arisen from recruiting some of the <a href="https://www.theguardian.com/science/2017/nov/02/big-tech-firms-google-ai-hiring-frenzy-brain-drain-uk-universities/">best machine learning and AI scientists</a>, organising them into goal-driven teams, and freeing them up from having to teach or apply for funding.</p>
<h2>Mind reader</h2>
<p>DeepMind appears to have learned from the Royal Free data breach, having <a href="https://deepmind.com/blog/ico-royal-free/">“reflected” on its own actions</a> when it was signed on to work with the trust. It said that the breast cancer dataset it will receive from Royal Surrey is “de-identified”, which should mean that patients’ personal identities won’t be shared. </p>
<p>Another key difference is that the Royal Surrey <a href="https://medphys.royalsurrey.nhs.uk/imagedb/">dataset was explicitly collected for research</a> – indicating that participants gave consent for their data to be shared in this way. DeepMind has also been upfront about its approach to data access, management and security. It has appointed <a href="https://deepmind.com/blog/trust-confidence-verifiable-data-audit/">independent reviewers and verifiable data audits</a>, in the hope of building <a href="https://deepmind.com/applied/deepmind-health/data-security/">trust and confidence</a>. </p>
<p>Given DeepMind’s continued collaboration with the NHS on a range of research, citizens are rightly concerned about how private corporations might exploit the data they have willingly shared for publicly funded work.</p>
<p>Few details about the Royal Surrey research project – which is in the early stages of development – have been released, but it’s likely that DeepMind will focus on applying deep neural networks for scanning mammogram images to automatically identify signatures of cancerous tissue. This approach would be similar to its <a href="https://deepmind.com/blog/announcing-deepmind-health-research-partnership-moorfields-eye-hospital/">Moorfields Eye Hospital</a> project, where DeepMind is building automated machine learning models that can predict macular degeneration and blindness from retinal scans. </p>
<p>DeepMind is probably also exploring the possibility of incorporating novel <a href="https://deepmind.com/research/publications/human-level-control-through-deep-reinforcement-learning/">deep reinforcement learning algorithms</a> to train the machine learning models. The algorithms would then tap into insights from <a href="http://www.cell.com/neuron/fulltext/S0896-6273(17)30509-3">empirical neuroscience</a> research about how the human brain learns from reward and punishment. </p>
<p><a href="https://mitpress.mit.edu/books/reinforcement-learning">Reinforcement learning</a> – which differs from more conventional <a href="https://machinelearningmastery.com/supervised-and-unsupervised-machine-learning-algorithms/">supervised and unsupervised learning methods</a> used in machine learning – is the technique that enabled DeepMind to train agents that learn to play Go and <a href="https://storage.googleapis.com/deepmind-media/dqn/DQNNaturePaper.pdf">many other games better than humans</a>. </p>
<p>Apart from these algorithmic advances, DeepMind might find, as it already has with the Streams trial, that there are many <a href="https://deepmind.com/blog/streams-and-ai/">technological tweaks that can improve how doctors treat patients</a>, without any need for machine learning at all.</p>
<h2>Human nature</h2>
<p>From my own experience in applying <a href="https://youtu.be/wTMsv5JWcEo">data analytics to medical diagnostics in neurology</a>, I know that – even if things go well for DeepMind and it manages to build a machine learning model that is excellent at detecting the early signs of breast cancer – it might well face a more practical problem in its application to the real world: <a href="https://www.darpa.mil/program/explainable-artificial-intelligence">interpretability</a>. </p>
<figure class="align-left ">
<img alt="" src="https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=237&fit=clip" srcset="https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/197953/original/file-20171206-943-15q17fe.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">‘Inflammatory’ breast cancer, in which tumour cells spread through dermal lymphatic channels (vascular invasion) of the skin.</span>
<span class="attribution"><a class="source" href="https://www.shutterstock.com/download/confirm/554909224?src=j42wLds0NclbTWp90FO9LQ-1-38&size=huge_jpg">Shutterstock</a></span>
</figcaption>
</figure>
<p>The practice of medicine today relies on trust between two humans: a patient and a doctor. The doctor judges the best course of treatment for a patient based on their individual clinical history, weighing up the relative pros and cons of the different options available. The patient implicitly trusts the doctor’s expertise. </p>
<p>But will patients or doctors trust a machine if it produced the same recommendation, based on an algorithm? While the use of machine learning has become commonplace in many contexts behind the scenes, <a href="https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/">its application in healthcare is fraught with the challenge of trust</a>. </p>
<p>If a doctor or patient fails to understand and communicate the rationale behind a recommendation, it might be very difficult to convince either to adopt it. And no machine learning algorithm is likely to be perfect. Both false positives and negatives are of great consequence in the healthcare context. </p>
<p>In the world of policing, AI has muddily kicked up hidden biases that can creep into machine learning models, <a href="https://techcrunch.com/2015/08/02/machine-learning-and-human-bias-an-uneasy-pair/">reflecting the unfairness embedded in our world</a>. Machines that inherit our prejudices might make for unpopular medical aids.</p>
<p>AI scientists are working on the problems of <a href="https://www.newscientist.com/article/mg23431195-300-bias-test-to-prevent-algorithms-discriminating-unfairly/">bias</a> and <a href="http://news.mit.edu/2017/using-machine-learning-improve-patient-care-0821">interpretability</a>, while also <a href="https://deepmind.com/blog/designing-with-clinicians/">working with clinicians</a> to design artificial intelligence <a href="https://wwwf.imperial.ac.uk/blog/explainable-ai/">that is more transparent about uncertainty</a>. </p>
<p>Beyond the technological advances in AI for improving human health, both ethics and interpretation will play central roles in its acceptance.</p><img src="https://counter.theconversation.com/content/88707/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Srivas Chennu receives funding from the UK Engineering and Physical Sciences Research Council and the Evelyn Trust. He has previously received funding from the National Institute for Health Research and the Royal Society.</span></em></p>DeepMind’s machine learning collaboration with another NHS trust (this time, it’s applying the tech to breast cancer) kicks up more questions of trust.Srivas Chennu, Lecturer in eHealth, University of KentLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/859792017-10-20T06:05:24Z2017-10-20T06:05:24ZGoogle’s new Go-playing AI learns fast, and even thrashed its former self<figure><img src="https://images.theconversation.com/files/191163/original/file-20171020-27065-1r5j84a.jpg?ixlib=rb-1.1.0&rect=0%2C670%2C6390%2C4119&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Better than human: the artificial intelligence that learned to master Go in just three days.</span> <span class="attribution"><span class="source">Shutterstock/maxuser</span></span></figcaption></figure><p>Just last year Google DeepMind’s <a href="https://deepmind.com/research/alphago/">AlphaGo</a> took the world of Artificial Intelligence (AI) by storm, showing that a <a href="https://theconversation.com/googles-go-victory-shows-ai-thinking-can-be-unpredictable-and-thats-a-concern-56209">computer program could beat the world’s best human Go players</a>. </p>
<p>But in a demonstration of the feverish rate of progress in modern AI, details of a new milestone reached by an improved version called <a href="https://deepmind.com/blog/alphago-zero-learning-scratch/">AlphaGo Zero</a> were <a href="https://www.nature.com/nature/journal/v550/n7676/full/nature24270.html">published this week in Nature</a>.</p>
<p>Using less computing power and only three days of training time, AlphaGo Zero beat the original AlphaGo in a 100-game match by 100 to 0. It wasn’t even worth humans showing up.</p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/why-google-wants-to-think-more-like-you-and-less-like-a-machine-79911">Why Google wants to think more like you and less like a machine</a>
</strong>
</em>
</p>
<hr>
<h2>Learning to play Go</h2>
<p>Go is a game of strategy between two players who take it in turns to place “stones” on a 19x19 board. The goal is to surround a larger area of the board than your opponent.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/191164/original/file-20171020-28465-1ljcwbo.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">The game of Go, simple to learn but a lifetime to master… for a human.</span>
<span class="attribution"><span class="source">Paragorn Dangsombroon/Shutterstock</span></span>
</figcaption>
</figure>
<p>Go has proved much more challenging than chess for computers to master. There are many more possible moves in each position in Go than chess, and many more possible games.</p>
<p>The <a href="https://blog.google/topics/machine-learning/alphago-machine-learning-game-go/">original AlphaGo first learned</a> from studying 30 million moves of expert human play. It then improved beyond human expertise by playing many games against itself, taking several months of computer time.</p>
<p>By contrast, AlphaGo Zero never saw humans play. Instead, it began by knowing only the rules of the game. From a relatively modest five million games of self-play, taking only three days on a smaller computer than the original AlphaGo, it then learned super-AlphaGo performance.</p>
<p>Fascinatingly, its learning roughly mimicked some of the stages through which humans progress as they master Go. AlphaGo Zero rapidly learned to reject naively short-term goals and developed more strategic thinking, generating many of the patterns of moves often used by top-level human experts. </p>
<p>But remarkably it then started rejecting some of these patterns in favour of new strategies never seen before in human play.</p>
<h2>Beyond human play</h2>
<p>AlphaGo Zero achieved this feat by approaching the problem differently from the original AlphaGo. Both versions use a combination of two of the most powerful algorithms currently fuelling AI: <a href="https://theconversation.com/no-more-playing-games-alphago-ai-to-tackle-some-real-world-challenges-78472">deep learning and reinforcement learning</a>.</p>
<p>To play a game like Go, there are two basic things the program needs to learn. The first is a policy: the probability of making each of the possible moves in a given position. The second is a value: the probability of winning from any given position.</p>
<p>In the pure reinforcement learning approach of AlphaGo Zero, the only information available to learn policies and values was for it to predict who might ultimately win. To make this prediction it used its current policy and values, but at the start these were random.</p>
<p>This is clearly a more challenging approach than the original AlphaGo, which used expert human moves to get a head-start on learning. But the earlier version learned policies and values with separate neural networks. </p>
<p>The algorithmic breakthrough in AlphaGo Zero was to figure out how these could be combined in just one network. This allowed the process of training by self-play to be greatly simplified, and made it feasible to start from a clean slate rather than first learning what expert humans would do.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/tXlM99xPQC8?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">How AlphaGo Zero learned to master Go.</span></figcaption>
</figure>
<p>An Elo rating is a widely used measure of the performance of players in games such as Go and chess. The best human player so far, <a href="https://www.goratings.org/en/players/1195.html">Ke Jie</a>, currently has an Elo rating of about 3,700.</p>
<p>AlphaGo Zero trained for three days and achieved an Elo rating of more than 4,000, while an expanded version of the same algorithm trained for 40 days and achieved almost 5,200.</p>
<p>This is an astonishingly large step up from the best human – far bigger than the current gap between the best human chess player <a href="https://ratings.fide.com/top_files.phtml?id=1503014">Magnus Carlsen</a> (about 2,800) and <a href="http://www.computerchess.org.uk/ccrl/4040/">chess program</a> (about 3,400).</p>
<h2>The next challenge</h2>
<p>AlphaGo Zero is an important step forward for AI because it demonstrates the feasibility of pure reinforcement learning, uncorrupted by any human guidance. This removes the need for lots of expert human knowledge to get started, which in some domains can be hard to obtain. </p>
<p>It also means the algorithm is free to develop completely new approaches that might have been much harder to find had it been been initially constrained to “think inside the human box”. Remarkably, this strategy also turns out to be more computationally efficient.</p>
<p>But Go is a tightly constrained game of perfect information, without the messiness of most real-world problems. Training AlphaGo Zero required the accurate simulation of millions of games, following the rules of Go. </p>
<p>For many practical problems such simulations are computationally unfeasible, or the rules themselves are less clear. </p>
<hr>
<p>
<em>
<strong>
Read more:
<a href="https://theconversation.com/no-more-playing-games-alphago-ai-to-tackle-some-real-world-challenges-78472">No more playing games: AlphaGo AI to tackle some real world challenges</a>
</strong>
</em>
</p>
<hr>
<p>There are still many further problems to be solved to create a general-purpose AI, one that can tackle a wide range of practical problems without domain-specific human intervention. </p>
<p>But even though humans have now comprehensively lost the battle with Go algorithms, luckily AI (unlike Go) is not a zero-sum game. Many of AlphaGo Zero’s games <a href="http://www.alphago-games.com/">have now been published</a>, providing a lifetime of inspirational study for human Go players.</p>
<p>More importantly, AlphaGo Zero represents a step towards a world where humans can harness powerful AIs to help find unimaginably (to humans) creative solutions to difficult problems. In the world of AI, there has never been a better time to Go for it.</p><img src="https://counter.theconversation.com/content/85979/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Geoff Goodhill does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>The new AlphaGo Zero artificial intelligence took just days to learn to play Go from scratch, with no human intervention. It even learned strategies never seen before in human play.Geoff Goodhill, Professor of Neuroscience and Mathematics, The University of QueenslandLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/784722017-06-05T20:02:23Z2017-06-05T20:02:23ZNo more playing games: AlphaGo AI to tackle some real world challenges<p>Humankind lost another important battle with artificial intelligence (AI) last month, when <a href="https://deepmind.com/research/alphago/">AlphaGo</a> beat the world’s leading Go player Ke Jie by three games to zero.</p>
<p>AlphaGo is an AI program developed by <a href="https://deepmind.com/">DeepMind</a>, part of Google’s parent company <a href="https://abc.xyz/">Alphabet</a>. Last year it <a href="http://www.abc.net.au/news/2016-03-15/google-ai-alphago-gets-divine-go-ranking/7249256">beat another leading player</a>, Lee Se-dol, by four games to one, but since then AlphaGo has substantially improved.</p>
<p>Ke Jie described AlphaGo’s skill as “<a href="https://www.cnet.com/au/news/google-alphago-ai-artificial-intelligence-go-ke-jie/">like a God of Go</a>”. </p>
<p>AlphaGo will now <a href="https://deepmind.com/blog/alphagos-next-move/">retire from playing Go</a>, leaving behind a legacy of games played against itself. They’ve been described by one Go expert as like “<a href="https://deepmind.com/research/alphago/alphago-vs-alphago-self-play-games/">games from far in the future</a>”, which humans will study for years to improve their own play.</p>
<h2>Ready, set, Go</h2>
<p>Go is an ancient game that essentially pits two players – one playing black pieces the other white – for dominance on board usually marked with 19 horizontal and 19 vertical lines.</p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=377&fit=crop&dpr=1 600w, https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=377&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=377&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=473&fit=crop&dpr=1 754w, https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=473&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/172208/original/file-20170605-31005-19hr6l0.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=473&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption">A typical game of Go: simple to learn but a lifetime to master.</span>
<span class="attribution"><a class="source" href="https://www.flickr.com/photos/alper/30626352/">Flickr/Alper Cugun</a>, <a class="license" href="http://creativecommons.org/licenses/by/4.0/">CC BY</a></span>
</figcaption>
</figure>
<p>Go is a far more difficult game for computers to play than chess, because the number of possible moves in each position is much larger. This makes searching many moves ahead – feasible for computers in chess – very difficult in Go.</p>
<p>DeepMind’s breakthrough was the development of general-purpose learning algorithms that can, in principle, be trained in more societal-relevant domains than Go.</p>
<p>DeepMind says the research team behind AplhaGo is <a href="https://deepmind.com/blog/alphagos-next-move/">looking to pursue other complex problems</a>, such as finding new cures for diseases, dramatically reducing energy consumption or inventing revolutionary new materials. It adds:</p>
<blockquote>
<p>If AI systems prove they are able to unearth significant new knowledge and strategies in these domains too, the breakthroughs could be truly remarkable. We can’t wait to see what comes next.</p>
</blockquote>
<p>This does open up many opportunities for the future, but challenges still remain.</p>
<h2>Neuroscience meets AI</h2>
<p>AlphaGo combines the two most powerful ideas about learning to emerge from the past few decades: deep learning and reinforcement learning. Remarkably, both were originally inspired by how biological brains learn from experience.</p>
<p>In the human brain, sensory information is processed in a series of layers. For instance, visual information is first transformed in the retina, then in the midbrain, and then through many different areas of the cerebral cortex.</p>
<p>This creates a hierarchy of representations where simple, local features are extracted first, and then more complex, global features are built from these.</p>
<p>The AI equivalent is called deep learning; deep because it involves many layers of processing in simple neuron-like computing units.</p>
<p>But to survive in the world, animals need to not only recognise sensory information, but also act on it. Generations of scientists and psychologists have studied how animals learn to take a series of actions that maximise their reward. </p>
<p>This has led to mathematical theories of reinforcement learning that can now be implemented in AI systems. The most powerful of these is temporal difference learning, which improves actions by maximising its expectation of future reward.</p>
<h2>The best moves</h2>
<p>By combining deep learning and reinforcement learning in a series of artificial neural networks, AlphaGo first learned human expert-level play in Go from 30 million moves from human games.</p>
<p>But then it started playing against itself, using the outcome of each game to relentlessly refine its decisions about the best move in each board position. A value network learned to predict the likely outcome given any position, while a policy network learned the best action to take in each situation. </p>
<p>Although it couldn’t sample every possible board position, AlphaGo’s neural networks extracted key ideas about strategies that work well in any position. It is these countless hours of self-play that led to AlphaGo’s improvement over the past year.</p>
<p>Unfortunately, as yet there is no known way to interrogate the network to directly read out what these key ideas are. Instead we can only study its games and hope to learn from these. </p>
<p>This is one of the problems with using such neural network algorithms to help make decisions in, for instance, the legal system: they can’t explain their reasoning. </p>
<p>We still understand relatively little about how biological brains actually learn, and neuroscience will continue to provide new inspiration for improvements in AI. </p>
<p>Humans can learn to become expert Go players based on far less experience than AlphaGo needed to reach that level, so there is clearly room for further developing the algorithms.</p>
<p>Also much of AlphaGo’s power is based on a technique called back-propagation learning that helps it correct errors. But the relationship between this and learning in real brains is still unclear.</p>
<h2>What’s next?</h2>
<p>The game of Go provided a nicely constrained development platform for optimising these learning algorithms. But many real world problems are messier than this, and have less opportunity for the equivalent of self-play (for instance self-driving cars).</p>
<p>So are there problems to which the current algorithms can be fairly immediately applied?</p>
<p>One example may be optimisation in controlled industrial settings. Here the goal is often to complete a complex series of tasks while satisfying multiple constraints and minimising cost.</p>
<p>As long as the possibilities can be accurately simulated, these algorithms can explore and learn from a vastly larger space of outcomes than will ever be possible for humans. Thus DeepMind’s bold claims seem likely to be realised, and as the company says, we can’t wait to see what comes next.</p><img src="https://counter.theconversation.com/content/78472/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Geoff Goodhill receives funding from the Australian Research Council and the National Health and Medical Research Council. </span></em></p>The artificial intelligence that beat a world master at the game of Go is now to be directed at more complex global problems. So what can we expect?Geoff Goodhill, Professor of Neuroscience and Mathematics, The University of QueenslandLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/784102017-05-26T13:05:17Z2017-05-26T13:05:17ZGoogle’s latest Go victory shows machines are no longer just learning, they’re teaching<p>Just over 20 years ago was the first time a <a href="https://theconversation.com/twenty-years-on-from-deep-blue-vs-kasparov-how-a-chess-match-started-the-big-data-revolution-76882">computer beat a human world champion</a> in a chess match, when IBM’s Deep Blue supercomputer beat Gary Kasparov in a narrow victory of 3½ games to 2½. Just under a decade later, machines were deemed to have conquered the game of chess when Deep Fritz, a piece of software running on a desktop PC, <a href="http://en.chessbase.com/post/kramnik-vs-deep-fritz-computer-wins-match-by-4-2">beat 2006 world champion Vladimir Kramnik</a>. Now the ability of computers to take on humanity has taken a step further by mastering the far more complex board game Go, with Google’s AlphaGo program <a href="https://www.nytimes.com/2017/05/25/business/google-alphago-defeats-go-ke-jie-again.html?_r=0">beating world number one</a> Ke Jie twice in a best-of-three series.</p>
<p>This signifcant milestone shows just how far computers have come in the past 20 years. DeepBlue’s victory at chess showed machines could rapidly process huge amounts of information, <a href="https://theconversation.com/twenty-years-on-from-deep-blue-vs-kasparov-how-a-chess-match-started-the-big-data-revolution-76882">paving the way for the big data revolution</a> we see today. But AlphaGo’s triumph represents the development of real artificial intelligence by a machine that can recognise patterns and learn the best way to respond to them. What’s more, it may signify a new evolution in AI, where computers not only learn how to beat us but can start to teach us as well.</p>
<p>Go is considered one of the <a href="https://www.quora.com/Is-Go-the-most-complicated-2-player-board-game">world’s most complex board games</a>. Like chess, it’s a game of strategy but it also has several key differences that make it much harder for a computer to play. The rules are relatively simple but the strategies involved to play the game are highly complex. It is also much harder to calculate the end position and winner in the game of Go. </p>
<p>It has a larger board (a 19x19 grid rather than an 8x8 one) and an unlimited number of pieces, so there are many more ways that the board can be arranged. Whereas chess pieces start in set positions and can each make a limited number of moves each turn, Go starts with a blank board and players can place a piece in any of the 361 free spaces. Each game takes on average twice as many turns as chess and there are six times as many legal move options per turn.</p>
<p>Each of these features means you can’t build a Go program using the same techniques as for chess machines. These tend to use a “brute force” approach of analysing the potential of large numbers of possible moves to select the best one. Feng-Hsiung Hsu, one of the key contributors to the DeepBlue team, argued in 2007 that <a href="http://spectrum.ieee.org/computing/software/cracking-go">applying this strategy to Go</a> would require a million-fold increase in processing speed over DeepBlue so a computer could analyse 100 trillion positions per second.</p>
<h2>Learning new moves</h2>
<p>The strategy used by AlphaGo’s creators at Google subsidiary DeepMind was to create an artificial intelligence program that could learn how to identify favourable moves from useless ones. This meant it wouldn’t have to analyse all the possible moves that could be made at each turn. In preparation for its first match against professional Go player Lee Sedol, AlphaGo analysed <a href="https://www.wired.com/2017/05/googles-alphago-levels-board-games-power-grids">around 300m moves</a> made by professional Go players. It then used what are called deep learning and reinforcement learning techniques to <a href="https://blog.google/topics/machine-learning/what-we-learned-in-seoul-with-alphago/">develop its own ability</a> to identify favourable moves.</p>
<p>But this wasn’t enough to enable AlphaGo to defeat highly ranked human players. The software was run on custom microchips specifically designed for machine learning, known as tensor processing units (TPUs), to support very large numbers of computations. This seems similar to the approach used by the designers of DeepBlue, who also developed custom chips for high-volume computation. The stark difference, however, is that DeepBlue’s chips could only be used for playing chess. AlphaGo’s chips run Google’s general-purpose AI framework, Tensorflow, and are also used to <a href="https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-with-custom-chip.html">power other Google services</a> such as Street View and optimisation tasks in the firm’s data centres.</p>
<h2>Lesson for us all</h2>
<p>The other thing that has changed since DeepBlue’s victory is the respect that humans have for their computer opponents. When playing chess computers, it was common for the human players to adopt so-called <a href="https://www.chess.com/blog/ramin18/anti-computer-tactics-gaming">anti-computer tactics</a>. This involves making conservative moves to prevent the computer from evaluating positions effectively.</p>
<p>In his first match against AlphaGo, however, Ke Jie, adopted tactics that had previously been used by his opponent to <a href="https://www.wired.com/2017/05/revamped-alphago-wins-first-game-chinese-go-grandmaster/">beat it at its own game</a>. Although this attempt failed, it demonstrates a change in approach for leading human players taking on computers. Instead of trying to stifle the machine, they have begun trying to learn from how it played in the past.</p>
<p>In fact, the machine has already influenced the professional game of Go, with grandmasters <a href="https://deepmind.com/blog/exploring-mysteries-alphago/">adopting AlphaGo’s strategy</a> during their tournament matches. This machine has taught humanity something new about a game it has been playing for over 2,500 years, liberating us from the experience of millennia.</p>
<p>What then might the future hold for the AI behind AlphaGo? The success of DeepBlue <a href="https://theconversation.com/twenty-years-on-from-deep-blue-vs-kasparov-how-a-chess-match-started-the-big-data-revolution-76882">triggered rapid developments</a> that have directly impacted the techniques applied in big data processing. The benefit of the technology used to implement AlphaGo is that it can already be applied to other problems that require pattern identification.</p>
<p>For example, the same techniques have been applied to <a href="https://www.wired.com/2017/05/using-ai-detect-cancer-not-just-cats/">the detection of cancer</a> and to create robots that can learn to do <a href="https://www.wired.com/2017/01/googles-go-playing-machine-opens-door-robots-learn/">things like open doors</a>, among <a href="https://www.wired.com/2017/01/googles-go-playing-machine-opens-door-robots-learn/">many other applications</a>. The underlying framework used in AlphaGo, Google’s TensorFlow, has been made freely available for developers and researchers to build new machine-learning programs using standard computer hardware. </p>
<p>More excitingly, combining it with the many computers available through the internet cloud creates the promise of delivering <a href="https://cloud.google.com/tpu/">machine-learning supercomputing</a>. When this technology matures then the potential will exist for the creation of self-taught machines in wide-ranging roles that can support complex decision-making tasks. Of course, what may be even more profound are the social impacts of having machines that not only teach themselves but teach us in the process.</p><img src="https://counter.theconversation.com/content/78410/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Mark Robert Anderson does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>Google’s AlphaGo victory over the human world champion shows how far things have come since DeepBlue.Mark Robert Anderson, Professor in Computing and Information Systems, Edge Hill UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/773832017-05-11T01:03:54Z2017-05-11T01:03:54ZComputers to humans: Shall we play a game?<figure><img src="https://images.theconversation.com/files/168795/original/file-20170510-21596-p2i8u6.png?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Artificial intelligence can bring many benefits to human gamers.</span> <span class="attribution"><a class="source" href="https://www.instagram.com/gamingartbysj/">Sam Jordan Belanger</a>, <a class="license" href="http://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND</a></span></figcaption></figure><p>Way back in the 1980s, a schoolteacher challenged me to write a computer program that played tic-tac-toe. I failed miserably. But just a couple of weeks ago, I explained to one of my computer science graduate students how to solve tic-tac-toe using the so-called “<a href="https://en.wikipedia.org/wiki/Minimax">Minimax algorithm</a>,” and it took us about an hour to write a program to do it. Certainly my coding skills have improved over the years, but computer science has come a long way too.</p>
<p>What seemed impossible just a couple of decades ago is startlingly easy today. In 1997, people were stunned when a chess-playing IBM computer named <a href="http://www.nytimes.com/1997/05/12/nyregion/swift-and-slashing-computer-topples-kasparov.html">Deep Blue beat international grandmaster Garry Kasparov</a> in a six-game match. In 2015, Google revealed that its DeepMind system had mastered several <a href="http://www.techrepublic.com/article/google-ai-beats-humans-at-more-classic-arcade-games-than-ever-before/">1980s-era video games</a>, including teaching itself a crucial winning strategy in “<a href="https://www.youtube.com/watch?v=V1eYniJ0Rnk">Breakout</a>.” In 2016, Google’s AlphaGo system beat a top-ranked Go player in a <a href="https://www.theatlantic.com/technology/archive/2016/03/the-invisible-opponent/475611/">five-game tournament</a>.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/V1eYniJ0Rnk?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">An artificial intelligence system learns to play ‘Breakout.’</span></figcaption>
</figure>
<p>The quest for technological systems that can beat humans at games continues. In late May, AlphaGo will take on <a href="https://arstechnica.com/information-technology/2017/04/deepmind-alphago-go-ke-jie-china/">Ke Jie</a>, the best player in the world, among other opponents at the Future of Go Summit in Wuzhen, China. With increasing computing power, and improved engineering, computers can beat humans even at games we thought relied on human intuition, wit, deception or bluffing – like <a href="http://www.csd.cs.cmu.edu/news/carnegie-mellon-ai-takes-chinese-poker-players">poker</a>. I recently saw a video in which volleyball players practice their serves and spikes against <a href="https://www.youtube.com/watch?v=EHKv6lRRV10">robot-controlled</a> rubber arms trying to block the shots. One lesson is clear: When machines play to win, human effort is futile. </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/EHKv6lRRV10?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">Robots play volleyball.</span></figcaption>
</figure>
<p>This can be great: We want a perfect AI to drive our cars, and a tireless system looking for signs of cancer in X-rays. But when it comes to play, we don’t want to lose. Fortunately, AI can make games more fun, and perhaps even endlessly enjoyable.</p>
<h2>Designing games that never get old</h2>
<p>Today’s game designers – who write releases that <a href="http://www.businessinsider.com/here-are-the-top-10-highest-grossing-video-games-of-all-time-2012-6">earn more than a blockbuster movie</a> – see a problem: Creating an unbeatable artificial intelligence system is pointless. Nobody wants to play a game they have no chance of winning.</p>
<p>But people do want to play <a href="https://theconversation.com/the-future-is-in-interactive-storytelling-76772">games that are immersive, complex and surprising</a>. Even today’s best games become stale after a person plays for a while. The ideal game will engage players by adapting and reacting in ways that keep the game interesting, maybe forever.</p>
<p>So when we’re designing artificial intelligence systems, we should look not to the triumphant Deep Blues and AlphaGos of the world, but rather to the overwhelming success of massively multiplayer online games like “<a href="https://worldofwarcraft.com/en-us/">World of Warcraft</a>.” These sorts of games are graphically well-designed, but their key attraction is interaction. </p>
<p>It seems as if most people are not drawn to extremely difficult logical puzzles like chess and Go, but rather to meaningful connections and communities. The real challenge with these massively multi-player online games is not whether they can be beaten by intelligence (human or artificial), but rather how to keep the experience of playing them fresh and new every time.</p>
<h2>Change by design</h2>
<p>At present, game environments allow people lots of possible interactions with other players. The roles in a dungeon <a href="https://en.wikipedia.org/wiki/Raid_(gaming)">raiding party</a> are well-defined: Fighters take the damage, healers help them recover from their injuries and the fragile wizards cast spells from afar. Or think of “<a href="https://en.wikipedia.org/wiki/Portal_2">Portal 2</a>,” a game focused entirely on collaborating robots puzzling their way through a maze of cognitive tests.</p>
<p>Exploring these worlds together allows you to form common memories with your friends. But any changes to these environments or the underlying plots have to be made by human designers and developers.</p>
<p>In the real world, changes happen naturally, without supervision, design or manual intervention. Players learn, and living things adapt. Some organisms even <a href="http://dx.doi.org/10.1086/691101">co-evolve</a>, reacting to each other’s developments. (A similar phenomenon happens in a <a href="http://www.amnh.org/exhibitions/einstein/peace-and-war/nuclear-arms-race/">weapons technology arms race</a>.)</p>
<p>Computer games today lack that level of sophistication. And for that reason, I don’t believe developing an artificial intelligence that can play modern games will meaningfully advance AI research. </p>
<h2>We crave evolution</h2>
<p>A game worth playing is a game that is unpredictable because it adapts, a game that is ever novel because novelty is created by playing the game. Future games need to evolve. Their characters shouldn’t just react; they need to explore and learn to exploit weaknesses or cooperate and collaborate. <a href="http://www.livescience.com/474-controversy-evolution-works.html">Darwinian evolution and learning</a>, we understand, are the drivers of all novelty on Earth. It could be what <a href="https://theconversation.com/evolving-our-way-to-artificial-intelligence-54100">drives change in virtual environments</a> as well.</p>
<p>Evolution figured out how to create <a href="https://theconversation.com/understanding-the-four-types-of-ai-from-reactive-robots-to-self-aware-beings-67616">natural intelligence</a>. Shouldn’t we, instead of trying to code our way to AI, just evolve AI instead? Several labs – <a href="http://hintzelab.msu.edu/">including my own</a> and that of <a href="http://adamilab.msu.edu/">my colleague Christoph Adami</a> – are working on what is called “<a href="https://en.wikipedia.org/wiki/Neuroevolution">neuro-evolution</a>.”</p>
<p>In a computer, we simulate complex environments, like a road network or a biological ecosystem. We create virtual creatures and challenge them to evolve over hundreds of thousands of simulated generations. Evolution itself then develops the best drivers, or the best organisms at adapting to the conditions – those are the ones that survive. </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/5lJuEW-5vr8?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">A neuro-evolution learns to drive a car.</span></figcaption>
</figure>
<p>Today’s AlphaGo is beginning this process, learning by continuously <a href="https://www.theguardian.com/technology/2016/jun/27/alphago-deepmind-ai-code-google">playing games against itself</a>, and by analyzing records of games played by top Go champions. But it does not learn while playing in the same way we do, experiencing unsupervised experimentation. And it doesn’t adapt to a particular opponent: For these computer players, the best move is the best move, regardless of an opponent’s style. </p>
<p>Programs that learn from experience are the next step in AI. They would make computer games much more interesting, and enable robots to not only function better in the real world, but to adapt to it on the fly.</p><img src="https://counter.theconversation.com/content/77383/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Arend Hintze receives funding from NSF BEACON Center for the Study of Evolution in Action Cooperative Agreement No. DBI-0939454, and received funding from Strength in Numbers Game Studio </span></em></p>Twenty years after Deep Blue beat Garry Kasparov at chess, artificial intelligence can make games more fun, and perhaps even endlessly enjoyable, if it learns to adapt.Arend Hintze, Assistant Professor of Integrative Biology & Computer Science and Engineering, Michigan State UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/622962016-07-13T13:03:25Z2016-07-13T13:03:25ZWhy football, not chess, is the true final frontier for robotic artificial intelligence<p>The perception of what artificial intelligence was capable of began to change when chess grand master and world champion <a href="http://www-03.ibm.com/ibm/history/ibm100/us/en/icons/deepblue/">Garry Kasparov lost to Deep Blue</a>, IBM’s chess-playing program, in 1997. Deep Blue, it was felt, had breached the domain of a cerebral activity considered the exclusive realm of human intellect. This was not because of something technologically new: in the end, chess was felled by the brute force of faster computers and clever heuristics. But if chess is considered the game of kings, then the east Asian board game Go is the game of emperors. </p>
<p>Significantly more complex, requiring even more strategic thinking, and featuring an intricate interweaving of tactical and strategical components, it posed an even greater challenge to artificial intelligence. Go relies much more on pattern recognition and subtle evaluation of the general positions of playing pieces. With a number of possible moves per turn an order of magnitude greater than chess, any algorithm trying to evaluate all possible future moves was expected to fail. </p>
<p>Until the early 2000s, programs playing Go progressed slowly, and could be beaten by amateurs. But this changed in 2006, with the introduction of two new techniques. First was the <a href="https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/">Monte Carlo tree search</a>, an algorithm that rather than attempting to examine all possible future moves instead tests a sparse selection of them, combining their value in a sophisticated way to get a better estimate of a move’s quality. The second was the (re)discovery of deep networks, a contemporary incarnation of neural networks that had been experimented with since the 1960s, but which was now cheaper, more powerful, and equipped with huge amounts of data with which to train the learning algorithms.</p>
<p>The combination of these techniques saw a drastic improvement in Go-playing programs, and ultimately <a href="https://theconversation.com/googles-go-triumph-is-a-milestone-for-artificial-intelligence-research-53762">Google DeepMind’s AlphaGo program beat Go world champion Lee Sedol</a> in March 2016. Now that Go has fallen, where do we go from here?</p>
<h2>The future of AI is in physical form</h2>
<p>Following Kasparov’s defeat in 1997, scientists considered that the challenge for AI was not to conquer some cerebral game. Rather, it needed to be physically embodied in the real world: football.</p>
<p>Football is easy for humans to pick up, but to have a humanoid robot running around a field on two legs, seeing and taking control of the ball, communicating under pressure with teammates, and all mostly without falling over, was considered completely out of the question in 1997. Only a handful of laboratories were able to design a walking humanoid robot. Led by <a href="http://sbiaustralia.org/systems-biology/kitanoprofile/">Hiroaki Kitano</a> and <a href="https://www.cmu.edu/me/people/veloso.html">Manuela Veloso</a>, the ambitious goal set that year was to have by 2050 a team of humanoid robots able to play a game of football against the world champion team according to FIFA rules, and win. And so the RoboCup competition was born.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/tAd1IeovyY8?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
</figure>
<p>The <a href="http://www.robocup.org/">RoboCup tournament</a> held its <a href="http://www.robocup2016.org/en/">20th competition in Leipzig this year</a>. Its goal has always been to improve and challenge the capacity of artificial intelligence and robotics, not in the abstract but in the much more challenging form of physical robots that act and interact with others in real time. In the years since, many other organisations have <a href="https://theconversation.com/cybathlon-will-showcase-what-bionics-could-do-for-millions-with-disabilities-54760">recognised how such competitions boost technological progress</a>.</p>
<p>The first RoboCup featured only wheeled robots and simulated 2D football leagues, but soon leagues that permitted Sony’s <a href="http://www.sony-aibo.com/">four-legged AIBO robot dogs</a> were introduced and, since 2003, <a href="http://wiki.robocup.org/wiki/Humanoid_League">humanoid leagues</a>. In the beginning, the humanoids’ game was quite limited, with very shaky robots attempting quivering steps, and where kicking the ball almost invariably caused the robot to fall. In recent years, their ability has significantly improved: many labs now boast <a href="http://robocup.herts.ac.uk/">five or six-a-side humanoid robot teams</a>.</p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/pzYHAp7b7sY?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
</figure>
<h2>No ordinary ballgame</h2>
<p>In order to push competitors on to reach the goal of a real football match by 2050, the conditions are made harder every year. Last year, the green carpet was replaced by artificial turf, and the goalposts and the ball coloured white. This makes it harder for robots to maintain stability and poses a challenge of recognising the goals and ball. So while the robots may seem less capable this year than the year before, it’s because the goalposts are moving.</p>
<p>The tasks involved in playing football, although much more intuitive to humans than chess or Go, are a major challenge for robots. Technical problems of hitherto unimaginable complexity have to be solved: timing a kick while running, identifying the ball against a glaring sun, running on wet grass, providing the robot with sufficient energy for 45 minutes’ play, even the materials that go into constructing a robot can’t disintegrate during a forceful game. Other problems to be solved will define important aspects of our life with robots in the future: when a robot collides with a human player, who can take how much damage? If humans commit fouls, may a robot foul back? </p>
<p>RoboCup offers up in miniature the problems we face as we head towards intelligent robots interacting with humans. It is not in the cerebral boardgames of chess or Go, but here on the pitch in the physical game of football that the frontline of life with intelligent robots is being carved out.</p><img src="https://counter.theconversation.com/content/62296/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Daniel Polani has been heading teams participating at the RoboCup competition since 1998. He was member of the executive, later the trustee board and is now president elect of the RoboCup Federation.</span></em></p>Computers must master football if they are to demonstrate that they can be our equal.Daniel Polani, Professor of Artificial Intelligence, University of HertfordshireLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/621812016-07-12T12:10:36Z2016-07-12T12:10:36ZCare.data has been scrapped, but your health data could still be shared<figure><img src="https://images.theconversation.com/files/129842/original/image-20160708-24101-2frk4u.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">
</span> <span class="attribution"><a class="source" href="http://www.shutterstock.com/cat.mhtml?lang=en&language=en&ref_site=photo&search_source=search_form&version=llv1&anyorall=all&safesearch=1&use_local_boost=1&autocomplete_id=&searchterm=patient%20records&show_color_wheel=1&orient=&commercial_ok=&media_type=images&search_cat=&searchtermx=&photographer_name=&people_gender=&people_age=&people_ethnicity=&people_number=&color=&page=1&inline=405862876">MangoNic/Shutterstock.com</a></span></figcaption></figure><p>Following a review by Dame Fiona Caldicott, the UK government decided to pull the plug on <a href="https://theconversation.com/nhs-care-data-still-leaks-like-a-sinking-ship-but-ministers-set-sail-regardless-43977">care.data</a>, a controversial NHS initiative to store all patient data on a single database. This may seem like a victory for data-privacy advocates, but NHS data-sharing initiatives are still being planned and the goalposts are being moved on patient consent.</p>
<p>A lot of NHS-related news was released on the day the <a href="http://www.iraqinquiry.org.uk/the-report/">Chilcot report</a> was published – presumably in the hope it would be buried. Alongside a crucial statement from health secretary Jeremy Hunt about the <a href="https://www.theguardian.com/society/2016/jul/06/jeremy-hunt-to-impose-new-contract-on-junior-doctors">junior doctors’ contracts</a>, there were the two related items of the Caldicott report on <a href="https://www.gov.uk/government/publications/review-of-data-security-consent-and-opt-outs">medical data and consent</a> and a <a href="https://www.gov.uk/government/speeches/review-of-health-and-care-data-security-and-consent">statement</a> by minister for life sciences, George Freeman, which announced the end of care.data. </p>
<p><a href="https://www.england.nhs.uk/ourwork/tsd/care-data/">Care.data</a> was an initiative that aimed to add patient records from GPs’ surgeries with existing data already collected by the NHS data centre, <a href="http://www.hscic.gov.uk/">HSCIC</a>. The resulting single database could then be used for medical research, NHS planning and maybe even <a href="http://www.hscic.gov.uk/media/12866/caredata-addendum-Information-Governance-Assessment/pdf/care.data_addendum_-_IG_assessment_-_September_2013_%20NIC-178106-MLSWX.A0913%20.pdf#10623256491942934180">commercial exploitation</a>. The <a href="http://www.legislation.gov.uk/ukpga/2012/7/contents/enacted">2012 Health and Social Care Act</a> ensured that any data sharing would be legal, even in the absence of patient consent.</p>
<p><a href="https://www.youtube.com/watch?v=Udpaajqg3nE&feature=youtu.be&t=14m20s">In April 2013</a>, Hunt promised that patients would be allowed to opt out of their GP data going to HSCIC (“type 1” objection), and any of their data leaving HSCIC (“type 2” objection). These options were communicated to the public in a <a href="https://www.england.nhs.uk/wp-content/uploads/2014/01/cd-leaflet-01-14.pdf">doormat leaflet</a>. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=454&fit=crop&dpr=1 600w, https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=454&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=454&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=571&fit=crop&dpr=1 754w, https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=571&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/130011/original/image-20160711-9302-5kbeq9.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=571&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Jeremy Hunt promised opt-outs.</span>
<span class="attribution"><a class="source" href="http://www.epa.eu/politics-photos/government-referenda-photos/cabinet-meeting-follwoing-uk-eu-referendum-photos-52855327">Andy Rain/EPA</a></span>
</figcaption>
</figure>
<p>But GPs raised concerns about the clarity of the information in the leaflet. They also raised concerns about patient confidentiality. To add to the anti-care.data sentiment, stories of earlier dubious NHS <a href="https://theconversation.com/time-for-some-truth-about-who-is-feeding-off-our-nhs-data-23998">data-sharing deals</a> began to emerge. These issues and others ensured ongoing postponements of the launch of care.data. </p>
<p>Type 1 opt-outs came into effect, not that it mattered as the care.data upload of GP data to HSCIC never happened. About 700,000 patients chose the type 2 opt-out but, even though the opt-out was applicable immediately, it was <a href="https://ico.org.uk/about-the-ico/news-and-events/news-and-blogs/2016/04/ico-statement-on-health-and-social-care-information-centre-undertaking/">disregarded by HSCIC</a> <a href="https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/517522/type2objections.pdf">until April 2016</a>.</p>
<h2>Caldicott review</h2>
<p>Caldicott was asked to review NHS data handling, including consent and opt-outs of care.data. Caldicott’s much delayed report said that the government should “consider the future of the care.data programme”, but it didn’t go as far as to say it should be axed. So how and why did the report lead to the care.data programme being scrapped?</p>
<p>The answer lies in the recommendations made by Caldicott on opt-outs. They back-pedal significantly on the concessions made by Hunt in 2013. The easiest way for the government to recant Hunt’s concessions is to withdraw the care.data programme they are associated with, and start with a clean slate. To be fair, the report strongly recommends an extended <a href="https://www.gov.uk/government/consultations/new-data-security-standards-for-health-and-social-care">public consultation</a>, so the new model isn’t a done deal yet. </p>
<p>In Caldicott’s new proposals, medical records from GPs’ surgeries can be sent to HSCIC without patients’ consent. The report argues that HSCIC is a safe haven for all medical data. </p>
<p>Significant limitations on type 2 objections have also been proposed. For “legally mandatory” data collection for HSCIC, such as Hospital Episode Statistics (data on treatment in NHS hospitals), opt-outs won’t apply. Also, opt-outs won’t apply to any anonymised data passed on to HSCIC’s customers, such as NHS divisions and research organisations.</p>
<p>The new Caldicott consent model explicitly excludes the use of medical data for <a href="https://ico.org.uk/action-weve-taken/enforcement/pharmacy2u-ltd">marketing</a> and <a href="http://www.telegraph.co.uk/news/health/news/10656893/Hospital-records-of-all-NHS-patients-sold-to-insurers.html">insurance</a> purposes. But companies that do data analysis for the NHS are viewed as “inside the tent” of a partially privatised NHS and will not need patient consent to receive data. </p>
<h2>Care.data is dead, long live care.data</h2>
<p>The man who officially pulled the plug on care.data, George Freeman, <a href="https://www.gov.uk/government/speeches/review-of-health-and-care-data-security-and-consent">makes it clear</a> that despite the end of care.data, medical data sharing is still firmly on the table, stating that “the government and the health and care system remain absolutely committed to realising the benefits of sharing information”. </p>
<p>This also shows in the NHS’s plans for care.data <a href="https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/442831/Work_Stream_2_2.pdf">successors</a>, one of which appears to be a “<a href="https://www.whatdotheyknow.com/request/single_national_gp_dataset_2">single GP dataset</a>” – basically, care.data without the opt-outs. </p>
<p>Caldicott and the government are taking a new negotiating position with GPs and patients on data sharing, in which the usefulness of the data takes the upper hand. The mechanism proposed for “consent” consists of providing information, and then granting patients a few limited opt-outs. As this does not sit well with the new European data protection law’s idea of consent (a “clear affirmative action”), we should expect further developments, both in the public debate and on the legal side.</p>
<p><em>Correction: On July 21, 2016, the sentence: “Type 1 opt-outs never came into effect as no data was sent from GPs’ surgeries to HSCIC.” was replaced with: “Type 1 opt-outs came into effect, not that it mattered as the care.data upload of GP data to HSCIC never happened.”</em></p><img src="https://counter.theconversation.com/content/62181/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Eerke Boiten receives funding from the UK government for the Kent Academic Centre of Excellence in Cyber Security Research, as well as from the EU for an Innovative Training Network in Cyber Security.</span></em></p>Data-privacy advocates may have won the care.data battle, but it looks like they’re about to lose the war.Eerke Boiten, Senior Lecturer, School of Computing and Director of Academic Centre of Excellence in Cyber Security Research, University of KentLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/589012016-05-05T13:00:15Z2016-05-05T13:00:15ZGoogle is now involved with healthcare data – is that a good thing?<figure><img src="https://images.theconversation.com/files/121378/original/image-20160505-29090-1qwo0s2.jpg?ixlib=rb-1.1.0&rect=0%2C748%2C4167%2C2950&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Your data is as important as who gets to see it.</span> <span class="attribution"><span class="source">phipatbig/shutterstock.com</span></span></figcaption></figure><p>Google has some of the most powerful computers and smartest algorithms in the world, has hired some of the best brains in computing, and through its purchase of British firm Deepmind has acquired AI expertise that recently <a href="https://theconversation.com/googles-go-triumph-is-a-milestone-for-artificial-intelligence-research-53762">saw an AI beat a human grandmaster at the game of go</a>. Why then would we not want to apply this to potentially solving medical problems – something Google’s <a href="https://theconversation.com/googles-larry-page-wants-to-save-100-000-lives-but-big-data-isnt-a-cure-all-28529">grandiose, even hyperbolic statements</a> suggest the company wishes to?</p>
<p>The New Scientist recently <a href="https://www.newscientist.com/article/2086454-revealed-google-ai-has-access-to-huge-haul-of-nhs-patient-data/">revealed</a> a <a href="https://drive.google.com/file/d/0BwQ4esYYFC04NFVTRW12TTFFRFE/view">data sharing agreement</a> between the Royal Free London NHS trust and Google Deepmind. The trust released incorrect statements (<a href="https://www.royalfree.nhs.uk/news-media/news/google-deepmind-qa/">since corrected</a>) claiming Deepmind would not receive any patient-identifiable data (it will), leading to irrelevant confusion about what data encryption and anonymisation can and cannot achieve.</p>
<p>As people have very strong feelings about third-party access to medical records, all of this has caused a bit of a scandal. But is this an overreaction, following <a href="https://theconversation.com/patients-will-resist-medical-record-sharing-if-nhs-bosses-ignore-their-privacy-fears-46147">previous health data debacles</a>? Or does this represent a new and worrying development in the sharing of medical records? </p>
<p><div data-react-class="Tweet" data-react-props="{"tweetId":"727890484044115968"}"></div></p>
<h2>Data analytics</h2>
<p>That the NHS outsources its data analysis requirements is nothing new. The NHS data centre <a href="http://www.hscic.gov.uk/">HSCIC</a> publishes regular data sharing reports, and <a href="http://www.hscic.gov.uk/media/20065/Data-Releases-Register-Oct---Dec-2015/xls/Copy_of_ReleaseRegister20151001_20151231_V03.xlsx">its latest report</a> details releases to companies such as CSL-UK, Northgate, McKinsey, and Dr Foster. These firms will sell the processed data back to the NHS.</p>
<p>Actually, while most NHS data sharing with companies is for so-called secondary purposes that lie outside the provision of direct clinical care, the deal with Google is classed as for direct care. Doctors get <a href="http://arstechnica.co.uk/business/2016/02/googles-deepmind-ai-group-working-with-nhs-to-develop-patient-care-software/">an app called Streams</a> which uses a patient’s live medical data and their historical record to determine their risk of acute kidney injury. </p>
<p>So it makes perfect sense for the app to access personally identifiable data of the patient being treated, and on that basis the claim that “<a href="http://www.dailymail.co.uk/news/article-3571433/Google-s-artificial-intelligence-access-private-medical-records-1-6million-NHS-patients-five-years-agreed-data-sharing-deal.html">Google has access to 1.6m patients’ data</a>” should not be cause for concern. Especially as Google accesses the data mostly indirectly, through an unnamed third party with certified information security standards, circumventing issues around potential abuse of the data by Google.</p>
<h2>Not so clear</h2>
<p>But another stated purpose of the deal is “real time clinical analytics, detection, diagnosis and decision support”, presumably with the intention of building an online platform for “medical-data-analysis-as-a-service”. Anything “as-a-service” normally implies the processing is done in the cloud, although the agreement with Google says little about that. Cloud processing means sensitive personal data will be sent to a Google server at some point.</p>
<p>The inclusion of five years of all patients’ historic data is justified to “aid service evaluation and audit of the new product”. But it’s hard to see how this is different from just using the data to improve the kidney injury algorithm in the first place. Deepmind’s <a href="https://deepmind.com/health">claim that “Streams does not use AI”</a> is downright bizarre in relation to the amount of data they claim to need, as this amount of data is usually used to feed machine learning algorithms that can then make better decisions because of it. Access to this trove of historic patient data will almost certainly come from Google itself.</p>
<p>Otherwise, the agreement with Google professes to be fully compliant with the Data Protection Act, standard medical data principles, and NHS procedures. Data transfer is secure (and encrypted), staff have been trained to respect confidentiality, and the data cannot be used for other purposes than those listed.</p>
<p>One principle mentioned is the <a href="https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/251750/9731-2901141-TSO-Caldicott-Government_Response_ACCESSIBLE.PDF">Caldicott principle</a> of using the minimum data required. But here this appears to be interpreted as: in order to treat one patient using Streams, we need five years’ medical data of 1.6m patients. This is seeing clinical care through a mass surveillance lens – we need all the data on everyone, just in case they require treatment. Conveniently, for clinical treatment matters NHS information governance allows the use of “implied consent” rather than any direct involvement from the subjects themselves.</p>
<h2>Black box surveillance</h2>
<p>The question is, of course, whether we trust Google to stick to these policies. The agreement allows for auditing by the NHS trust, and this may be enough of a deterrent against more direct and blatant abuses.</p>
<p>However, Google deals with personal data constantly: our search histories probably feed back on Google search rankings via some profiling process. Our Gmail emails are scanned for marketing purposes. If we stop Google from recording our location histories for our own use, do they still survive in the Google databases as some “anonymous” person’s location history? There is a lot here that Google is not telling us.</p>
<p>Improving the kidney injury algorithm or developing an analytics platform using medical data will generate more data. Service evaluation of the new product will generate more data. Some of that data will live in the shady world of people profiles, anonymised users, and aggregated user characteristics. It will be data that is somewhat personal but not personal enough for our crude data protection laws to be able to protect it. </p>
<p>In this world of <a href="http://www.hup.harvard.edu/catalog.php?isbn=9780674368279">black box surveillance</a>, Google is probably the world’s biggest player. As long as it offers so little transparency in how it uses and processes data, we have to be wary of it to some degree – and perhaps in this context specifically.</p><img src="https://counter.theconversation.com/content/58901/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Eerke Boiten receives funding from the UK government for the Kent Academic Centre of Excellence in Cyber Security Research, as well as from the EU for an Innovative Training Network in Cyber Security.</span></em></p>There are advantages, too.Eerke Boiten, Senior Lecturer, School of Computing and Director of Academic Centre of Excellence in Cyber Security Research, University of KentLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/563422016-03-20T19:27:58Z2016-03-20T19:27:58ZExplainer: Go and the ‘conversation of hands’<figure><img src="https://images.theconversation.com/files/115550/original/image-20160318-16336-1wl7ca9.png?ixlib=rb-1.1.0&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">Go is a beautiful and complex game that's endured for thousands of years. </span> <span class="attribution"><span class="source">Alexandre Keledjian</span>, <a class="license" href="http://creativecommons.org/licenses/by-nc/4.0/">CC BY-NC</a></span></figcaption></figure><p>Artificial intelligence reached a new frontier last week, when an AI defeated human Go champion Lee Se Dol <a href="https://gogameguru.com/alphago-defeats-lee-sedol-4-1/">four games to one</a>. </p>
<p>Google’s <a href="https://deepmind.com/alpha-go.html">Alpha Go</a> has made headlines for its ability to carry out the <a href="https://theconversation.com/ai-has-beaten-us-at-go-so-what-next-for-humanity-55945">complex calculations involved in the ancient Chinese game</a>, but I would like to give a different perspective. I want to talk about Go itself – an ancient game also known as baduk in Korean, weiqi in Chinese and Igo in Japanese - which ends, each time, with a beautiful representation of the player’s thoughts and strategies laid out across the board. </p>
<p>Go starts with an empty board of 19x19 squares. Two players take turns to place black or white stones anywhere on it, trying to surround a larger percentage of the board with their stones, or to limit the moves of the other player. </p>
<p>No stones are moved throughout the game, except when they are “captured,” by being surrounded. The aim of the game is to create spaces and connectedness. Go ends naturally, when both players agree there are no more useful moves to be made. </p>
<figure class="align-center zoomable">
<a href="https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=1000&fit=clip"><img alt="" src="https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=400&fit=crop&dpr=1 600w, https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=400&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=400&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=503&fit=crop&dpr=1 754w, https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=503&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/115564/original/image-20160318-16324-1hxv6b7.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=503&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px"></a>
<figcaption>
<span class="caption"></span>
<span class="attribution"><span class="source">2benny/Flickr</span>, <a class="license" href="http://creativecommons.org/licenses/by/4.0/">CC BY</a></span>
</figcaption>
</figure>
<p>The point of the game is not destroy your opponent, but to win with a small margin of points. It’s said that if a player is losing by more than eight points, they should resign. </p>
<p>Still, Go has never been about winning. Rather, it is about being able to develop oneself and learn. Perhaps this is why Go only made it to the West recently and instead chess, a game which is essentially destructive, has gained much more attention. </p>
<p>Go is equitable and deeply strategic, because each stone is equally valuable. The only thing that distinguishes a stone is the way it is placed at any given time. All have the potential to change the game.</p>
<p>Go derives from the Japanese word Igo. Although the game originated in China somewhere between three and five thousand years ago, it became known as Go during the <a href="http://www.britannica.com/event/Tokugawa-period">Edo period</a> (16th-19th centuries), when Japan established highly regarded and competitive schools and academies.</p>
<p>Although Japan has attracted and fostered world Go champions, such as the legendary father of the 20th century game, <a href="https://gogameguru.com/go-seigen/">Go Seigen</a>, Go has flourished throughout Asia. </p>
<p>The ethics of Go are deeply embedded in the Taoist and Confucius philosophies of self-mastery and the connection between humans and the natural environment. </p>
<p>Natural objects, such as stones and mountains, are attributed rights to exist regardless of the value they bring to the human sphere. Thus each tree or stone is intrinsically valued. </p>
<p>The four directions of the world symbolise the four sides of the Go board. The number of cross points on the board are equivalent to the number of lunar days in a year and the star points represent the most advantageous points on the board (the Goban).</p>
<figure class="align-right ">
<img alt="" src="https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=237&fit=clip" srcset="https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=600&fit=crop&dpr=1 600w, https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=600&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=600&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=754&fit=crop&dpr=1 754w, https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=754&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/115557/original/image-20160318-16319-1du4hf2.jpg?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=754&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Star points marked out on a standard Go board.</span>
<span class="attribution"><span class="source">Rommel2 via Wikimedia Commons</span></span>
</figcaption>
</figure>
<p>Go has been used to inform strategic decisions in governance and business. On a personal note, I used weekly games of Go to provide a conceptual framework for my PhD, by using the philosophical terms of space, connectedness and territoriality to describe the outcomes of civic engagement of recent migrants in Western Australia. The game has a supreme ability to challenge one intellectually, whilst remaining <a href="http://go-centre.nl/wp/forget-all-sorrows/">playful</a>. </p>
<p>The DeepMind challenge was not a competition, but a conversation between humans and non-humans. In the same way, Go is regarded as a <a href="http://search.proquest.com/openview/02116c59bc5b5ff32daf802d0dd6f07a/1?pq-origsite=gscholar">conversation of hands</a>. </p>
<p>When you ask for a game, you are asking “please teach me”. The player with opposite coloured stones is not only your opponent, but also your teacher and friend.</p>
<p>So let’s not forget in the debate about Google Artificial Intelligence that Go is foremost a game to be enjoyed. Much like life itself.</p><img src="https://counter.theconversation.com/content/56342/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Silvia Lozeva does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.</span></em></p>An artificial intelligence has defeated a world champion of Go, the ancient Chinese strategy game. But what is Go, and why is it worth teaching to a computer?Silvia Lozeva, Researcher and Lecturer , Curtin UniversityLicensed as Creative Commons – attribution, no derivatives.tag:theconversation.com,2011:article/539042016-03-17T10:07:19Z2016-03-17T10:07:19ZBeyond today’s crowdsourced science to tomorrow’s citizen science cyborgs<figure><img src="https://images.theconversation.com/files/115300/original/image-20160316-30247-1m7cr90.jpg?ixlib=rb-1.1.0&rect=863%2C450%2C4240%2C2345&q=45&auto=format&w=496&fit=clip" /><figcaption><span class="caption">All that computer power will still need a helping hand from our uniquely human expertise.</span> <span class="attribution"><a class="source" href="http://www.shutterstock.com/pic.mhtml?id=364388864">Computers image via www.shutterstock.com</a></span></figcaption></figure><p>Millions of citizen scientists have been flocking to projects that pool their time and brainpower to tackle big scientific problems, from astronomy to zoology. Projects such as those hosted by the <a href="https://zooniverse.org">Zooniverse</a> get people across the globe to donate some part of their cognitive surplus, pool it with others’ and apply it to scientific research.</p>
<p>But the way in which citizen scientists contribute to the scientific enterprise may be about to change radically: rather than trawling through mountains of data by themselves, they will teach computers how to analyze data. They will teach these intelligent machines how to act like a crowd of human beings.</p>
<p>We’re on the verge of a huge change – not just in how we do citizen science, but how we do science itself. </p>
<h2>The awesome human brain</h2>
<p>The human mind is pretty amazing. A young child can tell one human face from another without any trouble, yet it took computer scientists and engineers over a decade to build software that could do the same. And that’s not human beings’ only advantage: we are far more flexible than computers. Give a person some example images of galaxies instead of human faces, and she’ll soon outperform any computer running a neural net in classifying galaxies. </p>
<p>I hit on that reality when I was trying to classify about 50,000 galaxy images for my Ph.D. research in 2007. I took a brief overview of what computers could do and decided that none of the state-of-the-art solutions available was really good enough for what I wanted. So I went ahead and sorted nearly 50,000 galaxies “by eye.” This endeavor led to the Galaxy Zoo citizen science project, in which we invited the public to help astronomers classify a million galaxies by shape and discover the “weird things” out there that nobody knew are out there, such as <a href="https://en.wikipedia.org/wiki/Hanny%27s_Voorwerp">Hanny’s Voorwerp</a>, the giant glowing cloud of gas next to a massive galaxy. </p>
<h2>Enter the deep minds</h2>
<p>Computer scientists have made <a href="https://theconversation.com/evolving-our-way-to-artificial-intelligence-54100">significant steps forward in machine learning</a> over the last few years, and some of their inventions have started to <a href="https://www.reddit.com/r/MachineLearning/">hit the public consciousness</a>. </p>
<p>“Deep neural networks” (or deep minds) learn in a way that is closer to how our brains learn. They try to model data – say, photos of people – by turning them into high-level abstractions using multiple layers where each different layer may focus on different tasks (hence the “deep” in deep neural nets).</p>
<p>These machines are learning in a way that is more akin to what humans do: they start to develop their own intuition. Games are where computer scientists put their machine lab rats to the test. With their learned artificial intelligence, machines are starting to intuit which moves in a game are better than others. This is fundamentally different from previous approaches where the computers would try to “brute force” the game by calculating as many moves as possible and use smart statistics to figure out the best move that way. </p>
<figure>
<iframe width="440" height="260" src="https://www.youtube.com/embed/ePv0Fs9cGgU?wmode=transparent&start=0" frameborder="0" allowfullscreen=""></iframe>
<figcaption><span class="caption">Watch a deep mind totally ace ‘Space Invaders’</span></figcaption>
</figure>
<p>After just four hours of game play, a deep neural net developed by Google’s DeepMind has managed to come up with a Space Invaders strategy so optimal that <a href="http://www.wired.co.uk/magazine/archive/2015/07/features/deepmind">it was better than any person’s strategy</a>. </p>
<p><div data-react-class="Tweet" data-react-props="{"tweetId":"709670715390164992"}"></div></p>
<p>Recently, the team behind Google’s DeepMind has <a href="http://www.theguardian.com/technology/2016/feb/22/google-deepmind-go-alphago">thrown down the gauntlet</a> to the world’s best Go players, claiming that their deep mind can beat them. Go has remained an intractable challenge to computers, with good human players still routinely beating the most powerful computers – until now. Just this March AlphaGo, Google’s Go-playing deep mind, <a href="http://www.theguardian.com/technology/2016/mar/15/googles-alphago-seals-4-1-victory-over-grandmaster-lee-sedol">beat Go champion Lee Sedol 4-1</a>.</p>
<h2>Do we still need citizen science?</h2>
<p>We’re now entering an era in which machines are starting to become competitive with humans in terms of <a href="https://www.tensorflow.org/versions/r0.7/tutorials/image_recognition/index.html">analyzing</a> <a href="http://googleresearch.blogspot.ch/2015/06/inceptionism-going-deeper-into-neural.html">images</a>, a task previously reserved for human citizen scientists clicking away at <a href="http://www.galaxyzoo.org">galaxies</a>, <a href="http://www.oldweather.org">climate records</a> or <a href="http://www.snapshotserengeti.org/">snapshots from the Serengeti</a>. This landscape is completely different from when I was a graduate student just a decade ago – then, the machines just weren’t quite up to scratch in many cases. Now they’re <a href="https://theconversation.com/how-computers-help-biologists-crack-lifes-secrets-48416">starting to outperform people</a> in more and more tasks.</p>
<p>Rather than replacing citizen scientists, though, machines can help them – and it could not have come at a better time. Scientific experiments are <a href="https://theconversation.com/how-computers-broke-science-and-what-we-can-do-to-fix-it-49938">flooding researchers with data</a>: astronomers needed the help of the Internet to classify one million galaxies from an <a href="http://www.sdss.org/">astronomical survey </a>that took place in the 1990s and 2000s. Soon telescopes like the <a href="http://www.lsst.org/">Large Synoptic Sky Telescope</a> will give us images of billions of galaxies in addition to supernovae, asteroids and other strange things that go bump in the night. </p>
<p>How will astronomers be able to deal with all these data, many of which are time-sensitive? After all, if something goes “bump” and fades quickly, we’d want to try to study it more before it disappears forever. That’s where the machines can really help us: deep minds can scale up to process large data sets if we just give them sufficient processing power and memory. </p>
<h2>Citizen science cyborgs</h2>
<p>But the machines still need help – our help! One of the biggest problems for deep neural nets is that they require large training sets, examples of data (say, images of galaxies) which have already been carefully and accurately classified. This is one way in which the citizen scientists will be able to contribute: train the machines by providing high-quality training sets so the machines can then go off and deal with the rest of the data.</p>
<p>There’s another way citizen scientists will be able to pitch in: by helping us identify the weird things out there we don’t know about yet, the proverbial Rumsfeldian “unknown unknowns.” Machines can struggle with noticing unusual or unexpected things, whereas humans excel at it. </p>
<figure class="align-center ">
<img alt="" src="https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&fit=clip" srcset="https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=45&auto=format&w=600&h=427&fit=crop&dpr=1 600w, https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=30&auto=format&w=600&h=427&fit=crop&dpr=2 1200w, https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=15&auto=format&w=600&h=427&fit=crop&dpr=3 1800w, https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=45&auto=format&w=754&h=536&fit=crop&dpr=1 754w, https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=30&auto=format&w=754&h=536&fit=crop&dpr=2 1508w, https://images.theconversation.com/files/112963/original/image-20160225-15170-66vagj.png?ixlib=rb-1.1.0&q=15&auto=format&w=754&h=536&fit=crop&dpr=3 2262w" sizes="(min-width: 1466px) 754px, (max-width: 599px) 100vw, (min-width: 600px) 600px, 237px">
<figcaption>
<span class="caption">Hanny’s Voorwerp appears as a blue blob next to a galaxy in this screenshot of the Galaxy Zoo interface. Human citizen scientist Hanny van Arkel noticed it right away, while machines and algorithms trawling through the SDSS survey data did not.</span>
<span class="attribution"><span class="source">Galaxy Zoo/SDSS/K. Schawinski</span></span>
</figcaption>
</figure>
<p>Having the citizen scientists help the machines spot these unexpected things in the data would complement the machines’ ability to churn through huge data sets. If a machine got confused by something, or just wanted some extra feedback, it could kick the object back to a human for help, and then update itself to deal with similar things in the future. This could find applications not just in astrophysics, but in many other fields of science, from surveys of the sea floor to archives in museums, and the detectors of particle accelerators.</p>
<p>So envision a future where a smart system for analyzing large data sets diverts some small percentage of the data to human citizen scientists to help train the machines. The machines then go through the data, occasionally spinning off some more objects to the humans to improve machine performance as time goes on. If the machines then encounter something odd or unexpected, they pass it on to the citizen scientists for evaluation.</p>
<p>Thus, humans and machines will form a true collaboration: citizen science cyborgs.</p><img src="https://counter.theconversation.com/content/53904/count.gif" alt="The Conversation" width="1" height="1" />
<p class="fine-print"><em><span>Kevin Schawinski receives funding from the Swiss National Science Foundation. He is the co-founder of the Galaxy Zoo.</span></em></p>Computers are getting better and better at the jobs that previously made sense for researchers to outsource to citizen scientists. But don’t worry: there’s still a role for people in these projects.Kevin Schawinski, Assistant Professor of Galaxy & Black Hole Astrophysics, Swiss Federal Institute of Technology ZurichLicensed as Creative Commons – attribution, no derivatives.