Dynamics of Analog Electronic Neural Networks.

摘要

We review recent results on dynamics and stability of analog neural networks and discuss their application to associative memory and visual processing. Stability criteria for these networks, gaurantee convergence to fixed-point attractors under continuous-time and discrete-time, parallel updating. For associative memory, phase diagrams describing different attractor types are discussed, and it is shown that reducing analog transfer function steepness improves network performance. For visual processing, a two-dimensional, translation-invariant network is described. The network detects image features using a novel architecture that greatly reduces network wiring. Neural networks, Associative memory, Feature detection, Image processing, Analog computation.