Polanyi’s Paradox…how AI learnt human judgement

In the world of Artificial Intelligence (AI), why is it so challenging replicating common sense, intuition, judgement…all those mystical human abilities that we know are there but can’t quite define?

What most definitions propose is that these skills relate to a human’s ability to reason beyond data and calculation – we regularly make decisions based on factors other than statistics, and this skill isn’t congruent with the way we usually think about computers.

Mathematician Michael Polanyi studied the causes behind our ability to acquire knowledge that we can’t precisely explain.  Try describing the colour blue to someone – we know it when we see it but it’s not that easy to put into words.  Polanyi’s Paradox summarises the cognitive phenomenon that many times “we know more than we can tell”.

Polanyi’s Paradox has deep implications in the AI field: if we can’t explain our knowledge, how can we possibly train AI agents?

Thanks to Google, AI has evolved past Polanyi’s Paradox.  Take the game GO, for example.  Many of the well-accepted strategies in the ancient game are very hard to model as a series of rules and are typically more related to humankind’s intuition.  But then, between 2016 and 2017, DeepMind’s AlphaGo program regularly defeated the world’s top GO players.

AlphaGo overcame the Polanyi Paradox using some clever AI techniques.  Instead of relying on traditional symbolist’s algorithms such as inverse deduction to teach AlphaGo the rules of Go strategies, the DeepMind team used a combination of deep learning and reinforcement learning to train AlphaGo.  Initially, AlphaGo studied millions of Go games to infer hidden patterns between a specific strategy and the outcome of the game.  After that, researchers had AlphaGo play numerous games against itself to build new knowledge.

The AlphaGo example demonstrates that the way to build human-type judgment into AI systems is to design systems that learn on their own and include judgment-based decisions in the training data.

Makes one wonder about the implications for actuarial judgement…

Pamela Hellig