Implementation of non-holonomic ICA using sub-threshold analog CMOS circuits with automatic offset compensation

Abstract

An analog neurochip for nonholonomic independent component analysis (ICA) is de-signed under subthreshold operation. In order to incorporate the offsets due to device mismatches, modified algorithm using 2-quadrant multipliers and conventional biases is pro-posed. The proposed methods enhance the capability of the neural network strongly even in the presence of device mismatches. Analog approach which uses a massively parallel collective processing is best for implementation of neuromorphic system. But the crucial disadvantage is that analog circuitries are very susceptible to process variations. So some analog VLSI implementation for ICA, which is a highly attractive algorithm for signal processing, shows just preliminary results only for toy problems because they are easily affected by convergence problems, restricted dynamic ranges, and offset effects. So a new implementation for ICA, less subject to fabrication variation, has significance. For that, we propose two offset- tolerant enhancement learning methods. The one is to use a 2-quadrant multiplier for computing of updated weight values because it is more robust to backward-offsets than a 4-quadrant multiplier. The second is to compensate forward-offsets by conventional biases. Although the biases can not compensate the offsets at each iteration, in an expectation sense they can successfully compensate the offsets. The modified nonholonomic ICA algorithm, which uses 2-quadrant multipliers and extra biases, shows remarkable enhancements in both of convergence and offset tolerance. The proposed methods are very simple and automatically adjustable, so they are very appropriate to implementations of analog neural networks. The fabricated chips are tested with speech and sinusoidal mixtures. Test results for two women``s voices demonstrate that the neurochip successfully can separate the mixture by more than 10dB in a view of SNR. For a sub-Gaussian problem, it can enhance spectral powers of original signals ...