ABSTRACT Sensors are the universal interface of intelligent systems, yet they generate large volumes and mostly unnecessary data streams, much of it irrelevant to the eventual classification. Conventional architectures transmit these data at full resolution to multilayer perceptron and convolutional transformer classifiers. These are accurate, but they require task‐specific training and are memory‐traffic‐intensive, resulting in substantial bandwidth, latency, and energy challenges before digital inference. This motivates a universal near‐sensor computing framework. To address this, we introduce the electrical parallel‐resistive input summation module (E‐PRISM), a memristive near‐sensor architecture that performs analog subset coding, composing many inputs into a single readout with 2 10 resolvable states in a single parallel measurement. This universal and scalable primitive reduces raw sensor bandwidth by nearly an order of magnitude, lowers per‐frame energy by ∼20 ×, and decreases latency by more than two orders of magnitude relative to MLP pipelines. As a stand‐alone classifier, E‐PRISM is effective or low‐dimensional streams, achieving ∼95% on noisy 10‐bit pattern recognition, ∼88% on 2D shape classification, and ∼99% on motion‐trajectory tasks without digital post‐processing. For higher‐dimensional inputs, coupling E‐PRISM to a lightweight Kolmogorov–Arnold network yields a hybrid that consistently exceeds 95% accuracy across domains including 3D object recognition, wavelength discrimination, sensor fusion, and M‐of‐N safety logic.
Kumar et al. (Wed,) studied this question.