Brain-Like AI Chips Now Talk 3x Less — And Get Smarter

Most AI today runs on power-hungry chips in big data centres. But a growing branch of research copies how actual brains work: instead of crunching numbers constantly, these systems fire tiny electrical pulses, called spikes, only when something happens. The catch has always been that these spikes get passed along everywhere, even when the message is boring and predictable — like a security guard radioing in "all clear" every ten seconds. That chatter wastes energy, especially when devices have to send it wirelessly.

A new paper from researchers Aidin Attar and Michele Rossi fixes this by borrowing another trick from biology: prediction. Their design adds a small guessing layer between each stage of the network. It constantly predicts what the next signal will be. The system then only forwards the parts it got wrong — the genuinely surprising bits — and quietly suppresses everything it already expected. Think of a weather forecaster who only bothers telling you when the forecast changes.

The results are striking. On two standard tests — one recognising handwritten digits captured by a camera that senses movement, another recognising spoken digits — the new layers cut communicated activity by three times. Unexpectedly, accuracy went up too, not down. The authors suggest that by filtering out predictable noise, the network ends up with cleaner, more useful information at each stage. Less chatter, better signal.

Why does this matter beyond a research paper? Because the biggest limit on smart devices isn't computing power — it's battery and bandwidth. Wireless sensors in farms, hospitals, factories and homes spend most of their energy just transmitting data. Cutting that traffic by two-thirds could mean devices that need charging far less often, or that work in places where you can't easily plug anything in. It's early-stage lab work, so don't expect a product next month — but it points clearly at where efficient AI is heading.