Towards Analog VLSI Arrays for Nonseparable 3D Spatiotemporal Filtering

We report the development of a new CNN based theoretical framework and its on-going analog VLSI implementation for continuous-time discrete-space spatiotemporal filtering. Our proposed formulation, which stems from the theory of linear cellular neural networks, is capable of realizing a wide range of linear spatiotemporal dynamics. Based on our proposed framework, 3D nonseparable spatiotemporal filters are synthesized. The filters are presented in the form of continuous-time/discrete-space mixed domain transfer functions with responses qualitatively resembling visual cortical bandpass spatiotemporal receptive fields. Analog VLSI building blocks suitable for the implementation of the required spatiotemporal dynamics are presented. Simulation results on AMS 0.35mum process parameters of a 5times5 array, with electrical parameters tuned to realize the spatiotemporal dynamics of one of our synthesized 3D bandpass filters, are presented