Associative content-addressable networks with exponentially many robust stable states

• DOI: 10.48550/arxiv.1704.02019
• Publication Date: 2017-01-01
• Authors: Rishidev Chaudhuri; Ila Fiete
• Keywords: Associative property; Content (measure theory)

The brain must robustly store a large number of memories, corresponding to the many events encountered over lifetime. However, memory states in existing neural network models either grows weakly with size or recall fails catastrophically vanishingly little noise. We construct an associative content-addressable exponentially stable and robust error-correction. possesses expander graph connectivity on restricted Boltzmann machine architecture. expansion property allows simple dynamics perform at par modern error-correcting codes. Appropriate networks can be constructed sparse random connections, glomerular nodes, learning using low dynamic-range weights. Thus, quasi-random structures---characteristic important codes---may provide for high-performance computation artificial brain.