this is a little out there, but has anyone considered using deep neural networks to try representing programs visually? for example I could image a setup like an autoencoder that inputs and outputs the AST of a program, but the latent space is an image vector. given a large enough dataset, this would be a way to see if it's possible to come up with a universal graphical representation of all programs. I did something like this for representing unique letters/binary strings and am curious if I could extend it to "programs": https://github.com/noahtren/GlyphNet

Reminds me of Facebook's work on using neural networks to transform mathematical equations (https://ai.facebook.com/blog/using-neural-networks-to-solve-advanced-mathematics-equations/). Not so different from an AST. In fact, it's almost a special case of an AST, although the differences may turn out to be important.

@Konrad Hinsen thanks for sharing! Their results are impressive, I bet I could learn a lot by seeing how they represent mathematical expressions. I was thinking I might start with arithmetic or binary logic, anyways