Design and Simulation of Analog Vlsi Multi-Layer Perception Using Back-Propagation Learning

Abstract

Summary In the artificial neural network multi-layer perceptron is the most useN model. It has been successfully applied to solve problems in environment. However the slow learning procedures and complex computations of back propagation learning rules [l] prohibit this software methodology. In this work, a modular design for MLP is inaroduced. The design bases on current-mode subthreshold CMOS circuit. h4LP consists two modules, one is synapse which does weight multiplication and learning. and neuron which performs sigmoid transformation and error propagation. They are shown in Fig. 1. Fig. 2 shows the MLP architectws of two adjacent layers and their connections. The structure of MLP is extendible and suited to multi-chip modular design. In these modules, analog multipliers, sigmoid functions and weight storage elements are needed. Subthreshold CMOS technology owns the advantages of the implementation of neural system [2-31 for high integration density, low power dissipation, and similar BJT characteristics. Translinear application f4] is suited to the subthreshold CMOS circuit. Therefm, computationally power current-mode circuits can be synthesized. Fig. 3 shows a onequadrant curtent mode analog multiplier based on the translinear principle. Using differential Circuit technique, a four-quadrant multiplier is achieved and shown in Fig. 4.The SPICE simulation result is shown in Fig. 5. The sigmoid function can be achieved by a differentid pair. The complete circuit is shown in Fig. 6. Fig. 7 shows the simulation msult. Weight is stored in capacitor in voltage form, then feeds to a V/r converter. The learning circuit is shown in Fig. 8. From the simulation results, we find that the current mode analog multiplier on the mslinear principle is a very prospective method to implement the network multi-layer perceptxun using the back-propagation.