Convexity, internal representations and the statistical mechanics of neural networks

We present an approach to the statistical mechanics of feedforward neural networks which is based on counting realizable internal representations by utilizing convexity properties of the weight space. For a toy model, our method yields storage capacities based on an annealed approximation, which are in close agreement with one-step replica symmetry-breaking results obtained from a standard approach. For a single-layer perceptron, a combinatorial result for the number of realizable output combinations is recovered and generalized to fixed stabilities.