Silicon implementation of an auto-adaptive network for real-time separation of independent signals

A nonlinear autoadaptive filter for the separation of independent signal sources is described. The filter takes the form of a recurrent network which uses N simple processing units interconnected by N(N-1) inhibitory synapses. Each synapse determines its weight from a local Hebbian-like learning rule. Criteria are specified which the learning rule must satisfy, and examples from testing with digital simulations are given. The network was implemented using analog VLSI technology so as to achieve real-time, low-power, scalable solutions. Test results from a 2×_{chip are presented}