Oriented Spatial Pattern Formation in a Four Layer CMOS Cellular Neural Network

Cellular neural networks have found applications both as a platform for high speed image processing, as well as a model for complex spatiotemporal dynamical phenomena. This paper describes spatial pattern formation in a CMOS VLSI implementation of a CNN. With four layers of 32 × 64 cells, this CNN contains the most recurrently interconnected 2D layers of any VLSI CNN implementation to date. With over 8000 cells, it is also one of the largest in terms of the total number of cells. The spatial patterns generated by this network are bands of activity with a preferred scale and orientation, both of which can be adjusted via external bias voltages. The VLSI implementation provides a physical substrate for experimenting with complex spatiotemporal phenomena in real time, while also being amenable to theoretical analysis. Our experimental characterization of the observed patterns are in excellent agreement with our theoretical predictions.