3-Layer CNN Chip for Focal-Plane Complex Dynamics with Adaptive Image Capture

Abstract

This paper presents a CMOS implementation of a layered CNN concurrent with 32times32 photosensors with locally programmable integration time for adaptive image capture. The network is arranged in two layers containing feedback and control templates, inter-layer connections and programmable ratio of time constants. There are also feedforward connections to a third layer, which is faster, and devoted exclusively for combining the outputs of the other two. A more robust and linear multiplier block has been employed to reduce irregular analog wave propagation ought to asymmetric synapses. Global and local adaptation circuits are included on-chip. The predicted computing power per power consumption, 240MOPS/mW, is amongst the largest reported, what renders this kind of devices as especially adequate for portable applications of artificial vision