A Stochastic Memristive Photodetector for Bayesian Neuromorphic Vision

ABSTRACT Human vision achieves strong perception by performing probabilistic inference directly at the sensory level, integrating noise, prior knowledge and confidence in a unified process. Modern vision systems, however, remain dominated by camera‐centric, deterministic architectures that are energy‐intensive, latency‐limited and intrinsically unable to represent uncertainty at the point of sensing, creating a fundamental bottleneck for safe and efficient machine vision. Here, we demonstrate a stochastic memristive photodetector that combines unipolar resistive‐switching–driven randomness with optically induced, multilevel memory across two distinct time scales from milliseconds to seconds. These intrinsic device properties produce input‐sequence–dependent photocurrent outputs with probabilistic, multilevel statistics that naturally serve as feature encodings for a Bayesian computing framework. We implement and validate key probabilistic operations, including bit‐pattern detection, temporal pattern classification and discrimination of normal vs. abnormal heartbeat dynamics. We further simulate a device‐calibrated 64×64 Bayesian retina in which dual‐timescale pixels perform local Bayesian updates to convert dynamic scenes into in‐sensor spatiotemporal probability maps, and we demonstrate generality by achieving ∼79% tri‐class accuracy on chest x‐ray datasets (normal, Covid, pneumonia). Together, these results establish stochastic, memory‐bearing photodetectors as practical front‐end generators of probabilistic representations for Bayesian computing, opening a route toward uncertainty‐aware, energy‐efficient vision hardware.