Machine Learning learns Anew
Machine Learning algorithms trained by us mimic us well, but they never produce insights. Only occasionally, we come across work where the ML systems generate insights. One recent example was Alpha Zero learning chess by playing against itself with an occasional playoff against Stockfish. Stockfish is a great chess program that has been engineered over 10 years. But humans have taught it chess and thus its moves, while flawless, are always human like. Alpha Zero did not have any human training, it learned by on its own, and some of its moves are beyond creative. They are magical.
In another example of AI creativity, researchers form DeepMind find the principle components of a dataset by using ideas from Game Theory. The paper and the blog post are must reads for the details about the two key ideas the authors propose. First, the authors reformulate the age-old problem of finding Principal Components Analysis (PCA) to one finding the Nash Equilibrium of an appropriate multi-agent game. Second, they train the model using an approach similar to the one used by Alpha Zero. They use Reinforcement Leaning policies so that ‘EigenGame’ can discover the appropriate algorithm to compute the PCA by playing the game.
In EigenGame we see AI creating and not just mimicking. If approaches like this can generate insights about fundamental algorithms such as PCA, we can have have similar approaches deployed to many other interesting problems. Can an algorithm analyse all the astronomical observations and come up with new ideas that supersede Newton’s law of gravity and Einstein’s general relativity? The technology for this is not ready right now, but algorithms can definitely point us towards new and creative ideas. Alpha Zero is already changing chess.
