New Brain-Like Chip Learns On Its Own, Could Slash AI Power Bills
A chip that learns while it runs could mean far less power-hungry AI.
Most AI today works like a student who studies hard, then freezes. The "learning" happens on powerful servers, and the finished model is shipped off to your phone or laptop where it can only repeat what it already knows. A team of researchers in Spain has published a design for something different: a small analog circuit that keeps learning while it operates, the way your brain does when you meet new people or practice a skill. The paper was accepted at an IEEE circuits conference.
The trick is a component called a memristor — think of it as a resistor with a memory, a part whose electrical resistance changes depending on the current that has passed through it, and then stays changed. The researchers paired one with a small local circuit that watches the timing of incoming and outgoing electrical spikes (the brain's signals). If one spike comes just before another, the connection strengthens. If the order is reversed, it weakens. This is how real neurons learn, and here it happens automatically, with no external computer choreographing the process.
Their simulations, based on a 130-nanometer chip manufacturing process, showed the connections gradually adapting as the network ran. A tiny four-neuron test network even organized itself into specialized roles without anyone labeling the data — the electronic equivalent of a baby learning to tell faces apart without being told which is which.
Why should you care? Because the biggest obstacle to AI in everyday objects is power. Big models need data centers and huge electricity bills. Brain-style chips that learn locally could put smart sensing into hearing aids, factory sensors, pacemakers and doorbell cameras that run for months on a coin battery — and keep your data on the device instead of shipping it to a server. The honest catch: this is a simulation of a two-by-two neuron network on an older manufacturing node. A chip you can buy is likely years away, and scaling to millions of neurons is a genuinely hard problem.
- The circuit mimics how brain cells learn: connections get stronger or weaker based on the timing of electrical signals, with no outside software in charge.
- It relies on a memristor — a component that remembers how much current has flowed through it, acting like a tiny adjustable switch.
- The demo was small: a 2x2 neuron network simulated on an older 130nm chip design, so this is a research milestone, not a product you can buy.
Why It Matters
Could someday put self-learning AI into tiny battery-powered devices, cutting energy costs and keeping your data local.