In Douglas Adams’ iconic series A hitchhiker’s guide to the galaxya super-intelligent species created a super-powerful computer called Deep Thought to answer the ultimate question – what is the meaning of life (and the universe and everything)? Life imitates art so it should come as no surprise that a team in London founded a venture in 2010 called (I’m sure intentionally) DeepMind. The company was acquired by Google in 2014.
The team’s most widely-publicized achievement was beating a professional Go player with their AlphaGo program, which many consider a superior accomplishment to DeepBlue’s victory over Garry Kasparov in chess – in part simply due to the larger size of the board (19×19 versus 8×8), making a brute-force state-space search challenge even more wildly impossible. AlphaGo is also different in using deep neural networks to guide play.
The approach starts with an expert-trained policy network which assesses best moves to play from any given position. This is complemented by a value network which assesses a score for the position after the move. AlphaGo then does something quite unique – it plays thousands of games against itself, from the current position, refining these two networks to finally decide which move to make next. This self-training capability, at each move in the game, is what enables it to manage this vast search space and make it competitive against world-class players.
AlphaGo is implemented in Google Cloud across distributed CPU and GPU platforms (GPUs/DSPs are good for neural nets) – no indication of specialized hardware. Hardware-assist would presumably be possible, though is not clearly necessary.
In a more practical vein, DeepMind as an organization also has a big focus on healthcare. You’re probably thinking “ah-ha – deep reasoning for diagnosis”. You would be wrong as far as I can tell, at least today (though I’m sure there are plans). They have two products, Streams and Hark. Streams enables streaming clinical data (e.g. lab results) directly to mobile platforms. Hark is an early-stage clinical task management app – from what I have seen a sort of calendar of next tasks for a nurse or doctor. While neither of these apps seems very revolutionary, medical professionals involved in the development say that simply streamlining information flow and making sure tasks are not dropped will significantly improve patient outcomes.
What I find quite interesting about both of these apps is that they were developed in close collaboration with doctors and nurses at the Royal Free Hospital and at St. Mary’s Hospital, both in London. This is the right way to develop any app – directly with the people who will use it. And having built a level of trust, I’m sure DeepMind will have a much easier time introducing diagnostic aids based on deep reasoning.
Finally, the DeepMind team has given a lot of thought to data privacy and security. They commit that for UK patients, data is stored only in the UK and never linked to Google accounts, products or services – which should create an interesting challenge for Google who presumably plan synergies in the acquisition. However this pans out, looks like Google is developing an interesting direction in e/mHealth.
You can learn more about DeepMind HERE.
Share this post via:
Build a 100% Python-based Design environment for Large SoC Designs