An extension of Hopfields model with adaptive threshold and inhibitory interactions yields a network capable of nearly optimal storage of patterns of low activity. A replica symmetric solution of the mean-field equations is presented for the noise-free case. The following properties are demonstrated: For a low level of activity a the storage capacity increases as -(a lna)-1; up to 0.38 bits per synapse can be stored; spurious states can be suppressed; the network is not opinionated, i.e., it can categorize inputs as not similar enough to patterns stored.
ASJC Scopus subject areas
- Atomic and Molecular Physics, and Optics