Higher order memories in optimally structured neural networks

Based on one-step Hebbian learning we incorporate higher order correlations into the flexible architecture of a neural network optimized in first order. Sparse connectivity as well as a suitable summation technique eliminate the proliferation problem of higher order terms.