Extracting Formulae in Many-Valued Logic from Deep Neural Networks
This provides a method for improving interpretability of deep neural networks, which is an incremental advance in explainable AI.
The authors tackled the problem of interpreting deep ReLU networks by proposing an algorithm to extract formulae in many-valued logic from them, enabling logical formula extraction from trained networks with real-valued weights.
We propose a new perspective on deep ReLU networks, namely as circuit counterparts of Lukasiewicz infinite-valued logic -- a many-valued (MV) generalization of Boolean logic. An algorithm for extracting formulae in MV logic from deep ReLU networks is presented. As the algorithm applies to networks with general, in particular also real-valued, weights, it can be used to extract logical formulae from deep ReLU networks trained on data.