Brain-Like AI Gets Sharper When You Snip Most of Its Wiring

Researchers are trying to build AI that works more like a brain. So-called spiking neural networks only send a signal when a 'neuron' actually fires — think light switch, not a tap left dripping. That means far less energy burned, which matters for phones, smartwatches, hearing aids and medical implants. The catch: these brain-like systems are notoriously hard to train and hard to slim down for small devices.

A new paper asks a simple question: what happens if you delete the connections that barely fire? On ordinary AI, that trick works well. On brain-like AI, the usual tools — the ones designed for standard networks — fell apart, dropping to roughly coin-flip accuracy once about 20% of connections were removed. A different rule, which keeps busy connections and cuts quiet ones, held up instead.

Then came the surprise. When 80% of the connections were stripped from a network trained to recognise handwritten digits, accuracy didn't collapse. It rose slightly: 98.4% versus 97.2% for the original, untrimmed model. The author's guess is that those quiet connections mostly served other categories, adding background noise. Remove them, and the target category stands out more clearly — a little like tuning a radio by cutting the static.

Be careful with the hype. This is one author's simulation on a small, decades-old digit dataset, not a test on real chip hardware, and the gain is under two percentage points. The 'brain-like specialisation' explanation is offered as an analogy, not proof, and the author admits different settings might change the results. Still, the direction is promising: if brain-inspired AI can get smaller, cheaper and more personal without losing accuracy, it could quietly power the next wave of battery-limited gadgets.