Programmed interactions in higher-order neural networks: Maximal capacity (Q1179461)

The authors prove that in recurrent higher-order networks (in which every neuron produces the output \(x_{i_ 1}'\) given by the sign of its net income \(\sum w_{i_ 1i_ 2\ldots i_{d+1}}\cdot x_{i_ 2}\cdot x_{i_ 3}\cdots x_{i_{d+1}}\), for all \(1\leq i_ 2\leq i_ 3\ldots\leq i_{d+1}\)): {1. } The storage capacity of the outer- product algorithm generalized to networks of degree \(d\) is of the order of \(n^ d/\log n\) patterns. {2. } The maximal algorithm independent storage capacity is of the order \(n^ d/d!\). {3. } A spectral algorithm is presented for which near optimal capacity is conjectured.