Computer Science - Machine Learning

We describe a novel family of models of multi-layer feedforward neural networks in which the activation functions are encoded via penalties in the training problem. Our approach is based on representing a non-decreasing activation function as the …