Abstract
The static and dynamical properties of neural networks having many-neuron interactions are studied analytically and numerically. The storage capacity of such networks is found to be unchanged from that of the more widely studied case of two-neuron interactions implying that these networks store information no more efficiently. The size of the basins of attraction in the many-neuron case is calculated exactly from a solution of the network dynamics at full connectivity and reveals that networks with many-neuron interactions are better at pattern discrimination than the simpler networks with only two-neuron interactions