End-to-End Optimization of High-Density e-Skin Design: From Spiking Taxel Readout to Texture Classification

Abstract

Spiking readout architectures are a promising low-power solution for high-density e-skins. This paper proposes the end-to-end model-based optimization of a high-density neuromorphic e-skin solution, from the taxel readout to the texture classification. Architectural explorations include the spike coding and preprocessing, and the neural network used for classification. Simple rate coding preprocessing to spiking outputs from a modeled low-resolution on-chip spike encoder is demonstrated to achieve a comparable texture classification accuracy of 90 % at lower power consumption compared to the state of art. The modeling has also been extended from single-channel sensor recording to time-shifted multi-taxel readout. Applying this optimization to an actual tactile sensor array, the classification accuracy is boosted by 63 % for a low-cost FFNN using multi-taxel data. The proposed Spike-based SNR (SSNR) and Spike Time Error (STE) metrics for the taxel readout circuitry are shown to be good predictors of the accuracy.