A research team has built a spiking neural network that reads raw camera events one at a time, skipping the frame-by-frame processing that slows down most vision systems.
REACT, a fully spiking state-space model, processes each event as it arrives instead of grouping them into frames or time bins first. It relies on a new type of artificial neuron, called C-SiLIF, whose internal state updates based on the exact time gap between incoming events. Tested on gesture recognition and time-to-collision estimation, REACT hit a 9.59% relative error in collision prediction, within 0.15 percentage points of the best existing method, while responding in 4.6 milliseconds. At typical approach speeds, that latency window covers just 4 centimeters of vehicle movement, versus roughly 1 meter for the next-fastest learned method.
For anything moving fast, a robot dodging a ball, a car estimating time to impact, milliseconds translate directly into stopping distance. Shrinking that reaction gap from a meter of travel to four centimeters is the kind of unglamorous latency win that matters more than a flashy accuracy chart. The model's INT8 quantized version also cut estimated energy use from 18.5 to 2.8 millijoules per batch of events, a detail that matters for anything battery powered and constantly watching.
It will not replace conventional cameras in your phone, but for the narrow world of collision-avoidance sensors, being fast and cheap to run beats being marginally more accurate.