The Diffusion Network in Analog VLSI Exploiting Noise-Induced Stochastic Dynamics to Regenerate Various Continuous Paths

The Diffusion Network (DN) is a stochastic recurrent network capable of learning to regenerate various distributions of continuous-valued, continuous-time paths. By generalizing the data variability with internal stochasticity, the DN is found useful for distinguishing time-varying, biomedical signals reliably. In addition, the DN is a generalized form of the hidden Markov model and the Kalman filter, both of which have been useful in many applications. However, the continuous-time dynamics of the DN are governed by stochastic differential equations, making the DN unfavorable for execution in a digital computer. This paper presents the translation of the DN into analog very large scale integration (VLSI). By exploiting the natural differential current-voltage relationship of capacitors, the stochastic differential equations are computed by analog VLSI simultaneously in real time. The stochasticity required by the DN in VLSI is induced by a multi-channel noise generator on-chip, such that the DN actually uses noise-induced stochastic dynamics to generate continuous paths. Moreover, log-domain representation is employed to increase the dynamic range for diffusion processes, as well as to facilitate the subthreshold operation for power reduction. As subthreshold operation is prone to more nonlinear effects, the biasing conditions and operation ranges that minimize the effects are identified. Afterwards, a DN system-on-a-chip with four stochastic units is fabricated with the 0.18 μm CMOS technology. The measurement results reveal the DN in VLSI is able to regenerate a variety of continuous paths with noise-induced stochastic dynamics in VLSI. Moreover, the practical utility of the DN system is demonstrated in the context of recognizing electrocardiograms.