We investigate an approach for extracting knowledge from trained neural networks based on Angluin's exact learning model with membership and equivalence queries to an oracle. In this approach, the oracle is a trained neural network. We consider Angluin's classical algorithm for learning Horn theories and study the necessary changes to make it applicable to learn from neural networks. In particular, we have to consider that trained neural networks may not behave as Horn oracles, meaning that their underlying target theory may not be Horn. We propose a new algorithm that aims at extracting the ``tightest Horn approximation'' of the target theory and that is guaranteed to terminate in exponential time (in the worst case) and in polynomial time if the target has polynomially many non-Horn examples. To showcase the applicability of the approach, we perform experiments on pre-trained language models and extract rules that expose occupation-based gender biases.

翻译：暂无翻译