Scalable Models

The author of Gorgonia, a neural network library, had an interesting Strange Loop talk about reimplementing AlphaGo. The problem they ran into is that it would have cost them $70,000 to train a competitive model, so they focused on training their algorithm for games with a smaller board. What if a small neural network could be scaled to work on a larger board? The approach that I propose is based on program synthesis. Converting a neural network into a program has been demonstrated in this paper. A neural network trained on a smaller Go board could be converted into a program, and the program could then be modified to work on a larger Go board. A summary of program synthesis work can be found here.