Signed-weight projection for robust in-sensor computing neural classifiers under amplitude-dependent noise and stuck-at faults

In-sensor computing (ISC) monolithically integrates photodetection and neural-network inference on the same focal plane, avoiding analog-to-digital conversion and off-chip data movement. Despite its unrivalled latency and energy benefits, ISC remains limited by the fragility of analog weights. Amplitude-dependent drift, random-telegraph noise, and irreversible stuck-at faults can each trigger abrupt accuracy collapse. We introduce a training-free hardening strategy—signed-weight projection—that clips every weight to a symmetric range, enforcing a zero-mean distribution without changing network topology or requiring retraining. First-order perturbation analysis shows that the resulting balanced weights self-cancel the mean shift produced by multiplicative noise, while the magnitude cap limits the worst-case impact of saturated faults. Hardware measurements on a $$10\times 10$$ ISC prototype confirm a full optical-to-electrical inference path in $$\sim 2\,\mu \text {s}$$ , underscoring the need for on-array robustness. Device-level simulations on a (784–100–10) classifier further demonstrate graceful accuracy degradation. Noise tolerance and stuck-fault tolerance expand by well over an order of magnitude compared with an unprotected baseline, yet clean-data accuracy is preserved. The method offers a low-cost algorithm device co-design guideline, which can be retro-fitted to existing ISC pipelines and scaled to future high-resolution vision sensors.