Silicon experimentation of first order TDCNN dynjamics

Abstract

Recently, the authors have proposed a network formalism (TDCNNs) which introduces Time-Derivative coupling between linearized-CNN cells (with output nonlinearity removed) and demonstrated its use in realizing non-separable 3D spatiotemporal filters. TDCNNs assume inputs in the form of time-varying 2D array of pixels and processing is carried out in continuous-time. Due to this continuous-time nature of TDCNNs, it can be conveniently implemented with an array of continuous-time filters, each coupled to its nearest neighbors according to the feedforward/feedback and temporal-derivative templates. Analog circuit building blocks and simulation results from our first attempt in implementing TDCNNs with full custom CMOS was presented previously. This paper follows from our previous presentation and includes some of the measured results obtained from the fabricated prototype with 5 × 5 two-layered cells.