An implanted brain interface can't afford to ship every raw neural squiggle off-chip: the radio would drain the battery and swamp the bandwidth. So the sorting of spikes, deciding which neuron each blip came from, has to happen right there on the implant, under a brutal power budget. This paper's answer is a deep binarized neural network, whose weights are just plus or minus signs, so inference needs no multipliers, only sign flips and bit logic.

Working from tiny 16-sample spike snippets, it hit a median 98.7% accuracy on both synthetic and real recordings. The hardware numbers are the headline: an FPGA version decides each spike in about 0.01 milliseconds, and an ASIC estimate lands at 0.014 square millimeters of silicon drawing 122 nanowatts. That's a footprint that could actually sit inside the body.

Numbers are from the abstract, so check the paper for how the accuracy was measured across datasets.