A Tiny Memory Glitch Can Blind Car AI — This Fixes It

Self-driving cars, delivery drones and security cameras increasingly use "event cameras" — sensors that only record what changes in a scene, instead of filming every pixel all the time. That makes them fast and extremely power-thrifty. To read those images, engineers pair two kinds of AI: a brain-inspired "spiking" network that is very cheap to run, and a conventional AI network that is more accurate but burns more power. The trick is knowing when to hand off from the cheap one to the expensive one.

Until now, that handoff was designed around two things only: accuracy and battery drain. Nobody asked what happens when the hardware itself glitches — which it does. Chips in cars and drones get hit by heat, age and stray radiation, and a single memory bit can flip from 0 to 1. The BLADE team ran thousands of simulated faults and found something startling: one specific bit — the biggest number in a decimal calculation (the "exponent") — caused severe or catastrophic accuracy loss in 58.8% of the times it was flipped. Protecting that one bit costs roughly 3% more storage and removes those catastrophic failures entirely.

They also found a reassuring pattern: the more work you give the brain-like spiking network, the more resilient the whole system becomes. A version running entirely on spiking AI kept 96.5% of its reliability even under aggressive fault conditions. And when the conventional network exits early because it is confident, the whole detection step uses just 15.82 millijoules of energy — a tiny amount, well within what a small battery can supply.

The catch is that this is a research paper, not a product. It was tested on one event-based object detector under simulated faults, not on real chips in real cars over years of wear. Nothing here is shipping, and no company has announced plans to adopt it.