Learning from Each Other: Neural Networks and Finite Automata

The academic objective of this project is to explore how two well-known machine learning paradigms - finite automata and neural networks - can improve, explain and learn from each other. These two types of machine learning serve very different purposes and are developed in almost non-interacting communities. At the human level, this project aims to develop a network of researchers who can create flows of knowledge between these communities. This network will also investigate how these two types of machine learning can learn from each other at the algorithmic level. The title "Learning from each other" thus refers both to humans and algorithms. This project will initially consist of two tracks. The first will use finite automata to improve our understanding of neural networks. AI should be explainable, and this means understanding how neural networks solve complex problems by decomposing them into smaller ones whose solution is easier to interpret. Finite automata provide the perfect set of problems to investigate this question as they can be combined in logical ways to form arbitrarily complex problems. By systematically comparing how neural networks solve individual problems with how they solve combinations of problems we should uncover mechanisms of compositional reasoning in neural networks. The second track will use neural networks to improve the state-of-the-art algorithm for learning finite automata. This task has many industrial applications but faces a computational bottleneck consisting of generating certain counterexamples. We will train specialist neural networks called Large Language Models to generate counterexamples to solve this problem.