Brain-Like AI Gets a Power-Saving Trick That Cuts Energy Costs

Most AI you use today is powered by neural networks that crunch numbers in a giant, energy-hungry way. But there's a more efficient kind of AI called a spiking neural network, which works more like a human brain: neurons only 'fire' when they need to, like a light that only turns on when someone enters the room. These brain-like networks are great for saving power, but teaching them new tricks — a step called fine-tuning — has been painfully inefficient.

The problem is in the hardware. Fine-tuning normally requires a specialized chip called an in-memory computing accelerator, which runs calculations right where data is stored. But the standard way to fine-tune these spiking networks forces the chip to constantly rewrite its own data, like erasing and redrawing a whiteboard after every note. It also needs random number generators, which are bulky and drain power. Researchers from King's College London and partners found a way to skip those constant rewrites by cleverly combining the chip's normal calculations with a lightweight 'perturbation' signal. Their design, called IPZO, only activates when spikes fire, so it does the work without wasting energy on idle neurons.

The result? On standard tests, their method matched the accuracy of much more complex approaches while using significantly less power — in some cases, cutting the energy needed for fine-tuning by more than half compared to current methods. They also fixed a glitch that could distort results when reusing random numbers, using a simple XOR trick.

Why should you care? Because this could lead to AI that runs efficiently on your phone, smartwatch, or smart home devices without draining batteries or relying on the cloud. It also means data centers could shrink their massive electricity bills. The trade-off: the new approach requires a specially designed chip, so it may take a few years before it shows up in consumer gadgets. Still, it's a big step toward AI that's both smarter and greener.