SCAMP-5 Silicon Retina cuts event rate 47% while improving saliency

Dynamic vision sensors (DVS) mimic the retina's temporal contrast detection for high-speed, high-dynamic-range imaging, but they lack the spatial filtering and gain control that biological retinas use to distill information. In a new preprint on arXiv, Maciej Lewandowski and colleagues demonstrate the first programmable Silicon Retina that adds these bio-inspired stages directly onto the SCAMP-5 Pixel Processor Array—a massively parallel vision chip. They also built a GPU simulation framework to test the model on video intensity reconstruction and saliency prediction.

The results are striking: while the bio-inspired model is less effective at reconstructing absolute intensity frames, it beats standard DVS on the crucial task of saliency prediction—achieving a 13% lower loss using a lightweight FireNet-style network with only ~100k parameters. Even more impressive, the silicon retina's 'information distillation' reduces the event rate by about 47%, making the representation far more efficient for downstream neural networks. This suggests that adding a few biologically plausible processing stages can dramatically improve bandwidth-constrained edge AI applications, from autonomous drones to smart cameras.