Spiking Neural Networks Are Finally Doing Something Useful

So here's a question I've been asking myself for about a decade: when are spiking neural networks going to matter outside a university lab?

I remember sitting through a presentation at Automatica back in, I want to say 2014, watching a researcher explain how the brain uses spikes instead of continuous signals, and how this would revolutionize everything. My colleague Hans leaned over and whispered something unprintable about academic funding cycles. We both figured it was another five years away from being useful. Then five more years. Then five more.

Well, look, here's the thing. Two papers crossed my desk this week that have me reconsidering my skepticism. Not abandoning it entirely (I've been burned before), but reconsidering.

The first comes from a team working on autonomous vehicle perception. They've built a spiking neural network that processes LiDAR point clouds for object detection, the kind of thing every self-driving car needs to do constantly. The results are genuinely interesting: 92% average precision on easy detection scenarios, 87% on moderate, 86.5% on hard. Those numbers are competitive with conventional neural networks. But the real story is the energy consumption. They're claiming a 3.33x reduction in what they call "synaptic operation energy" compared to an equivalent CNN.

Now, I'll be honest, I had to call my old colleague at Siemens to sanity-check those numbers. His take: plausible, but the comparison methodology matters a lot. "Conservative loop-based operation" is doing some heavy lifting in that claim. Still, even if the real-world savings are half that, you're talking about meaningful power reduction for systems that need to run 24/7.

The second paper is the one that really got my attention, though. A team has built a flapping-wing micro aerial vehicle (basically a robotic butterfly, less than 30 grams) that uses spiking neural networks for fully onboard control. Not simulation. Not tethered flight. Actual autonomous flight with closed-loop control running on a $5 ESP32 microcontroller.

When I was at Kuka, we spent enormous effort on controller optimization for industrial arms. The compute requirements for real-time control are brutal, and that's for a stationary robot bolted to the floor. Doing it on a flying platform the weight of a few coins, with the kind of nonlinear dynamics you get from flapping wings... that's a different beast entirely.

The numbers they report: 36% latency reduction (1059 microseconds down to 680) and 18% power savings compared to conventional neural networks. The latency number is arguably more important than the power number for flight control. When your robot is constantly about to fall out of the sky, every microsecond counts.

What strikes me about both papers is that they're not using specialized neuromorphic hardware. The LiDAR work is designed for deployment on neuromorphic platforms, sure, but the flapping-wing robot runs on commodity hardware. A five-dollar chip. That's the kind of thing that actually scales.