Abstract

The notion of Boolean logic backpropagation was introduced to build neuralnetworks with weights and activations being Boolean numbers. Most ofcomputations can be done with Boolean logic instead of real arithmetic, bothduring training and inference phases. But the underlying discrete optimizationproblem is NP-hard, and the Boolean logic has no guarantee. In this work wepropose the first convergence analysis, under standard non-convex assumptions.

 

Quick Read (beta)