DeepMind’s deep learning software AlphaGo conquered one of the world’s greatest Go players, Lee Sedol, last week – a task that many thought would take another decade to accomplish. The AI approaches AlphaGo used to beat Sedol might be indicative of future trends in AI in general – particularly in the domain of healthcare.
In the world of one-to-one boardgames Go is special because, despite it’s apparent simplicity, the strategies required to play are considered to be intimately “human” in their reliance on intuition above calculation. While the combinations of many chess scenarios can sometimes be searched thoroughly to 3 or 4 moves head, the number of possible Go moves is borderline incalculable, so the same “brute force” computing strategies that allowed IMB’s Deep Blue to defeat Gary Kasparov are not possible on a Go board.
Where a human Go player can – in theory – play against herself, it’s most practical for her to play against another player to practice, train, and come in contact with different game scenarios. Meanwhile DeepMind had AlphaGo challenge various versions of itself millions of times over to help the system learn winning strategies.
In other words, AlphaGo was able to train to a degree that would be impossible for any human. Together, the policy network allows AlphaGo to determine the most promising maneuvers while the value network allows the system to work from a narrower data set and confidently make a decision without having to finish the entire match to confirm the decision. In human terms, policy networks and value networks give AlphaGo the statistical confidence to make quick and effective moves.
To date, the simplest breakdown of how AlphaGo searches an near-infinite number of moves to arrive at it’s best answer is in the video below by the publication Nature. At 3:10 into the video below, a description is presented of the “value networks” and “policy networks” that limit the searches to focus AlphaGo’s attention to the best possibilities on the board.
How might these novel approaches and developments in board game mastery help further the field of healthcare? Well, because DeepMind themselves are focusing more and more attention in that domain already.
DeepMind recently announced DeepMind Health and a partnership with the UK’s National Health Service (NHS). In a press release, DeepMind described how their mobile app will communicate information to nurses and doctors to help them detect acute kidney injury in their patients. A part of the DeepMind Health app called Hark is designed to to diminish healthcare givers’ reliance on beepers, fax machines, and mounds of paperwork by developing a clinical task management system.
Ultimately, systems such as these would allow healthcare services to detect and categorize illnesses incredibly fast through machine learning technology. Some of the opportunities are explored in this overview video about the NHS / DeepMind partnership.
AI giant IBM has put a lot of proverbial eggs in the basket of healthcare, allocating much of it’s Watson resources – in both development and marketing – to that field (see the Watson Health page here). With years of effort, the world hasn’t seen much of a tangible splash in healthcare from IBM’s efforts – might DeepMind’s novel approaches open up new options for AI health applications?
Here are some possibilities of how AlphaGo’s approaches might be leveraged in the domain of healthcare:
- An infectious disease detection program could “play against” versions of itself in a contest to correctly diagnose a series of infectious cases, improving itself based on which variation of the program correctly diagnosed the disease with the data provided (as with Go, this would involve a massive set of previously recorded disease cases)
- Like a Go board, the possibilities of where certain types of cancer (such as Leukemia) might spread are virtually limitless. It might be possible – through enough training data and computational ability – to train a computer to limit the reasonable search for a next spreading location, and preemptively treat regions most likely to be cancer targets
- Similarly, for diseases with a large potential scope of treatment options (combining medicines, dietary adjustments, etc…), a well-trained algorithm might be capable of limiting treatment options to combinations that are most likely to yield successful outcomes (rather than relying on a doctor’s preference alone, or on the treatment that seems most obvious initially)
If healthcare weren’t ripe with opportunity, we wouldn’t expect the best, brightest, and biggest (Google, IBM, and others) to be honed in on that sector. With DeepMind’s own health initiative and the inspirational victory of a novel combined approach to AI, there’s more firepower and focus on healthcare than ever.
Historically, IBM has won the fanfare game when it comes to AI. Winning in chess and Jeopardy! is part of the IBM claim to fame – but Go was a victory for DeepMind. There are too many players in the wide field of healthcare to tell who will claim a tangible AI breakthrough there, but whoever does will be playing for much more than board game bragging rights.
Image credit: Freepik.com