Artificial Intelligence and Medical Research: Time to Aim Higher?

I have the privilege to have been friends, since his Cambridge undergraduate days, with my newest hero Demis Hassabis. Demis, as I expect most readers will know by now, is the CEO and main public face of DeepMind, the artificial intelligence (AI) company that in 2014 was bought by Google for a cool $400 million, in 2015 developed a single program that mastered a slew of very diverse Atari video games just by being fed their responses to inputs and allowed to play them a lot, and in March 2016 used essentially the same underlying algorithm to play the board game Go better than the player who has dominated the game over the past decade. Demis does not just run DeepMind: he is its main founder and its intellectual inspiration, having forged a highly successful career starting in video games and moving to a PhD in neuroscience before combining those two fields in his current work.

Why is this such a big deal? Simplistically, the significance is that Go has always been the Holy Grail of computer game playing (a field with which I have some familiarity, since it was the focus of my very first program at age 15 and my undergraduate final-year project) because of its numerical complexity—there are typically about 200 legal moves in Go at any one time, as against about 20 in chess—and its subtlety, in which novices cannot look at a position and immediately say who is winning as they could in checkers just by counting pieces. But no, that is not the big deal, because complexity eventually falls to increasing computing power, and evaluation of the merit of a position can be hard coded by experts just as well when it is subtle as when it is not (though it would typically need more code). The big deal is that AlphaGo (that is the name of DeepMind's program) did not have anything hard coded into it concerning what makes this or that position better than another. It started out by knowing nothing but the rules of the game (which are extremely simple), then it studied a large database of games between reasonably skilled players, and then it played itself a huge number of times to refine what it considered a good move. This general idea has been attempted time and time again, but has always hit diminishing returns before achieving anything impressive. Deepmind's approach, which combines a couple of fashionable ideas in machine learning with a bunch of their own innovations, has not. And what makes this big deal even bigger is the speed with which AlphaGo got there: the project only began 2 years ago, and it reached good amateur level a year ago, middling pro level (top few hundred in the world) 6 months ago, and now #1.

I imagine by now you have a follow-up question, though: OK, so this is a really big deal in AI, but why is it an appropriate topic for an editorial in RR? The answer is not left to the imagination in Hassabis's comments to the press about DeepMind's plans for the future. Games, whether board or video, are mere testbeds for determining the versatility of algorithms such as DeepMind's machine learning approach: the actual goal is systems that can assist humanity in real-world tasks. Relatively primitive AI technology has already contributed to the performance of everything from web searching (such as for photos of a particular person) to self-driving cars (not so coincidentally two of Google's core interests), and we can be absolutely sure that DeepMind's findings will find their way into those applications in short order. The key difference when real-world applications are tackled is that there is no longer any scope for learning from unbounded quantities of data generated by self-play in isolation from the outside world—but the amount of data available anyway may very well suffice, finite though it may be.

So, what about antiaging medicine (another declared interest of Google, in the form of Calico)? Well, if we first look at medicine in general, the use of AI dates back decades, to the original so-called expert systems that attracted much interest in the 1970s and 1980s but have essentially failed to penetrate medical practice. The first problem they faced, as throughout medical innovation, is that any new technology must demonstrably improve on the pre-existing standard of care—and even if such an improvement exists, the more modest its magnitude the harder that “demonstrably” barrier is to overcome. The second is more psychological and is most prominently shared today with self-driving cars: even if improved performance is indeed objectively demonstrated, the public are still inclined to prefer to put their lives in the hands of a less reliable human than a more reliable machine.

It obviously remains to be seen whether these latest advances in machine learning will give computers a sufficient margin of performance over humans in the medical realm to tip the balance in terms of adoption. However, when we come to aging, there is another facet to this: the triple challenge that I discussed in this space a few years ago. ¹ Aging is generally, and rightly, considered the Holy Grail of medicine just as much as Go is the Holy Grail of board games. Although the complexity of aging is often overestimated, it is certainly very great—and the divide-and-conquer, damage-repair approach that I believe will ultimately defeat aging will entail the simultaneous administration of many therapies to people who are not yet experiencing declines in health, which in turn entails a huge requirement to anticipate and avert undesirable interactions. Demis and I agree that the understanding of aging and how to prevent it is probably too simple a problem to need the help of general-purpose AI. But the implementation of that understanding in the clinic, to the satisfaction of society, is a whole other challenge. That is, above all, why I view AlphaGo's achievement as a dramatic advance for medicine of the coming decades.