This paper describes an implementation of the Kurtosis and InfoMax algorithms for an independent components analysis in mixed-signal CMOS. Our design uses on-chip calibration techniques and local adaptation to compensate for the effect of device mismatch in arithmetic blocks and analog memory cells. We use our design to perform two-input blind source-separation on mixtures of audio signals and mixtures of EEG signals. Our experiments show that the hardware implementation of InfoMax consistently separates the signals within a normalized reconstruction error of less than 10%, while the reconstruction error of Kurtosis varies between 25% and 60%, due to its higher sensitivity to device mismatch and input statistics. Each circuit has a settling time of 8 ms, occupies a die area of 0.016-0.022 mm² and dissipates 15-20 mW of power.
An on-chip learning Artificial Neural Network (ANN) implementation
using the Pulse Width Modulatioii (PWM) technique is proposed in this paper. Synapse and neuron are analog circuits, while digital counters are utilized to store the weights. Through the PWM circuit, the digital weight is converted into a pulse signal as the input of the analog synapse circuit. The analog modified quantity of weight is transformed into a weight-update pulse signal whose width is proportion to the value of the weight modiřication quantity. The learning rule is bcised on the weight perturbation algorithrn. In this way, the weight can be long-terni-stored and ecisily modified, thereas the synapse and the neuron are of a small size in the silicon area and the learning Circuit is feasible for implementation. Taking the advantages of both the analog and the digital realizations of the ANN, this method is a meaningful way to the implementation of on-chip ANN and fuzzy processors.