neuroscience A field note by Vikrant Sharma
EPFL built a chip that watches two brain regions talk to each other
A new neural implant can decode real-time interactions between brain regions, not just record raw signals. That changes what closed-loop brain interfaces can do.
EPFL built a neural chip that does not just record brain activity. It decodes how two regions are talking to each other, in real time, on the chip itself. Most brain-computer interfaces collect raw voltage spikes from neurons and send them to a computer for analysis. This one runs the analysis locally. It listens to motor cortex and somatosensory cortex at the same time, figures out how they are coordinating, and adjusts stimulation on the fly. The chip computes spectral coherence between regions, which is a measure of how synchronised their oscillations are at different frequencies. The immediate use case is closed-loop deep brain stimulation for movement disorders. If the chip detects that motor planning signals are breaking down, it can trigger a corrective pulse without waiting for an external computer to decide. The latency drops from milliseconds to microseconds. What caught my attention is the on-chip compute. Previous attempts at real-time brain decoding either burned through power or needed offboard processing. EPFL’s chip runs at 1.72 milliwatts, low enough to sit inside a skull long-term. It uses a custom FFT engine and a parallel processing pipeline that handles 32 channels at once. The paper shows results from rat motor cortex during a reaching task. The chip correctly identified when the two regions were coupled during movement planning. That is not flashy, but it is hard to do in hardware without overheating or oversimplifying the signal. The part I want to know more about is how well this generalises. Spectral coherence is useful for movement disorders where oscillatory coupling matters. Does the same architecture work for memory tasks, where you care about spike timing or phase relationships across hippocampus and prefrontal cortex? Can it decode interactions between more than two regions without blowing the power budget? Still, this is the direction brain implants need to go. Offloading computation to the edge means faster feedback, lower power, and fewer cables sticking out of someone’s head.
Source: Neural chip decodes interactions between brain regions
