We are all AlphaGo Zero!
The DeepMind team's latest news on mastering the game of Go without human knowledge1 has sparked new interest in the future path of artificial intelligence (AI) development. AlphaGo Zero2, the new baby of DeepMind, defeated AlphaGo by huge margin, and practically any human player. Discussions about this new AI approach instantly filled university campuses as well as downtown cafés. The focus primarily was the fundamental shift in the AI paradigm from a data intensive methodology to one that requires no data at all. Some scientists have argued that intelligence is the result of learning from data, lots and lots of data! Data intensive methods are therefore the basis of artificial intelligence, regardless of the exact approaches adopted in training a machine to make certain decisions. Logic is almost never needed, but is a consolidation of experience learnt from data and training. That's precisely how we learned to do things. No kids ever learned to walk by following logic! Persistent experiments, repeated failures, and continuous improvements lead to mastery of a skill. Taking that to a high level, skills formulate logic, and logic creates rules, knowledge, even specific culture and ethics.
In a sense, AlphaGo Zero is doing exactly what most humans would do in learning to play any game. We had no data at the beginning, only the rule of the game. We kept playing and at the same time kept evaluating our own chance of winning. In the process, we generated our own data (or experience if you like), based on which we improved our skill and hence a better chance of winning. However, the key point is that everyone does this, but not everyone reaches the same level of proficiency. Some become very good but some remain very lousy players. So the question is: why such big differences? AlphaGo Zero is just one possible clever mind that trains itself with its own data so brilliantly. The good news is that there is no evidence suggesting such a clever mind is unique. So, there may still be many other ways!
November 2017
___________________
1David Silver et al., "Mastering the game of Go without human knowledge," Nature 500(19): 354–359, October 2017.
2AlphaGo Zero: Learning from Scratch