it's like a scene from Reservoir Dogs

#AutomatedTheoremProving #movies

Mathematics can be reduced to a few logical principles. In other words, Mathematics is a closed world problem.

This may mean that, like Chess or Go, a machine is better doing math
than human. Automatic theorem proving was extensibly studied and resulted, for example, fast SAT solvers but to achieve human-like creativity, deep learning based approaches might be a way to go.

There are three approaches I can think of
- GPT like model (fashionable)
- Reinforcement learning (there are some)
- Diffusion model (?)

Diffusion model would be interesting because a mathematician starts with an intuitive but wrong idea then polishes it to a rigorous proof.

I'm also interested in "bootstrapping", i.e. starting without any training data.

Any idea or reference?

#DeepLearning #AutomatedTheoremProving #TheoremProver