Differentiable Logic Cellular Automata

From Game of Life to pattern generation with learned recurrent circuits

by Pietro Miotti, Eyvind Niklasson, Ettore Randazzo, Alexander Mordvintsev

Imagine trying to reverse-engineer the complex, often unexpected patterns and behaviors that emerge from simple rules. (...) What if we could create systems that, given some complex desired pattern, can, in a fully differentiable fashion, learn the local rules that generate it, while preserving the inherent discrete nature of cellular automata?

This is really a nice work from the team at Google working on Paradigms of Intelligence. Mordvintsev was the inventor of DeepDream, that produced psychedelic images back in 2015.

What if we could directly learn these local rules, and create models that combine binary logic, the flexibility of neural networks, and the local processing of cellular automata?

What a nice question to have! I sense that this combination will in part afford explainability that today’s neural networks lack.

You recognize when something is going to be big when the authors explain it in such terms that you can follow their process even if you don’t understand the machinery behind their research. After reading this, I’m inclined to think this will have a powerful impact not only on image related tasks, but it will affect much broader fields.

Their work also points to robust computing, something that is of primary importance in a world of complex systems where modern complicated machines need to be resilient and robust. Think remote exploration (space, undersea, desert, whatever) where simple failures should not critical and result in full shutdown.

And a question, can something like DiffLogic CA, be used to learn the rules of a problem, but then be able to expand and improve? The effect observed of the diagonal construction of the checkerboard is a nice side effect. What would a DiffLogic CA capable of training on a dynamic problem do if it got that as input? These are completely speculative and in some sense absurd, or naive, but in any case I’ll be following this in the future.

read it in full at https://google-research.github.io/self-organising-systems/difflogic-ca/