Discretizing continuous neural networks using a polarization learning rule

Discrete neural networks are simpler than their continuous counterparts, can obtain more stable solutions, and their hidden layer representations are easier to interpret. This paper presents a polarization learning rule for discretizing multi-layer neural networks with continuous activation functions. This rule forces the activation value of a neuron towards the two poles of its activation function. First, we use this rule in the form of a modified error function to discretize the hidden units of a back-propagation network. Then, we apply the same principle to the second-order recurrent networks to solve grammatical inference problems. The experimental results are superior to the ones using existing approaches.