uSense: Unary-Computing-based Stochastic Edge Neuromorphic Sensing

Tactile sensing on the edge is a new frontier of embodied artificial intelligence, advancing novel applications like robotics, prostheses, and immersive virtual reality. However, the high-dimensional data generated from tactile arrays requires heavy computation and communication for real-time responses. To address this problem, we propose unary-computing-based stochastic edge neuromorphic sensing (uSense), a fast Fourier transform (FFT) framework based on unary computing for near-sensor tactile sensing. uSense encodes raw signals from the tactile sensors into rate-coded probabilistic bitstreams and performs stochastic FFT with extremely simple digital logic, alleviating memory, energy, bandwidth, and latency bottlenecks. Such benefits result from three key techniques: dual-domain (i.e., unary and binary) dynamic scaling, stage-wise bitwidth optimization, and computational graph pruning. First, dual-domain dynamic scaling ensures that our unary computing approach minimizes the numerical accuracy loss throughout the FFT computation. Second, stage-wise bitwidth optimization progressively tunes the data precision at each FFT stage, achieving one order of magnitude data compression with minimized accuracy loss. Third, computational graph pruning considers the task sensitivity of frequency components and removes redundant computation and memory access by 50%, achieving another order of magnitude data compression with minimized accuracy loss. Our experiments demonstrate that the uSense framework achieves 99.18% data compression, preserves 97.64% texture recognition accuracy across 21 textures, only 1.35% accuracy loss compared to floating-point baselines. Our analysis reveals an overlooked fact: high numerical inaccuracy does not necessarily translate to low decoding inaccuracy, as long as the structure of the feature latent space is preserved. Further analytical modeling on FFT hardware shows uSense can reduce the computation and memory footprint of FFT by more than 50%, and enable real-time responses of FFT in less than 100 nanoseconds.