Thursday, December 6, 2018

GammaGo

News (1) that AlphaGo can successfully learn Chess, Shogi and Go through self-play is interesting. It is symptomatic of trends in AI largely relying on raw computing power. Typically, innovation lags when resources become infinite and we have early signs of trouble here. Reinforcement learning through self-play is not a new concept - it has been here from the advent of computers. It is just that not many have access to computing resources necessary to create demonstrable prototypes.

More importantly, this approach is unlikely to culminate in cognition and consciousness, the possible end game. It is clearly the case that computers can create usable heuristics by repeated experiments, just as humans do. However, those heuristics are generated within a framework of rules that were specified ex. ante. The "deep mind," enthusiasts had argued a few years ago that their computer found a "new way," to play an ancient game. It is quite possible that given a large number of experiments, computers can learn from cases that are outside the norm. But to label this "creativity," is a stretch. It is more an accident than invention. One could argue that humans have benefitted handsomely from accidents in the past and so why not computers. This is true and so the general question is whether computing resources running amuck with an infinite number of repeated experiments can provide learnings from accidents at a faster rate than humans are capable of.

It is tantalizing. What the AI leaders need to understand, however, is that we have been here before. A critical look at the approach may be beneficial. We knew that we could predict from historical data ever since math was invented and we knew that repeated search of the design space could yield usable results since the advent of computers. The question is whether we have done anything new except pouring money into scaling conventional technologies. Stacking countless "computers," in the "cloud" on the promise of AI has many drawbacks.

It is time to go back to the drawing board. A field replete with engineers seems to be going in the wrong direction. As innovation lags in materials and quantum processing, they are creating mountains of Silicon to show heuristics generation is possible. The mathematicians locked up in low productivity areas such as finance, may be well advised to go back and think.

Thinking has a low premium currently and that is problematic.


(1) http://science.sciencemag.org/content/362/6419/1140