Spiking Neural Networks Are Having a Moment, and This Time It Might Actually Matter

The 3.33x energy reduction comes from a block-wise analysis of synaptic operations, and yes, that's under "conservative loop-based operation" assumptions, so your mileage may vary on actual neuromorphic hardware. But even if you cut that number in half, you're still looking at meaningful efficiency gains for a task that runs continuously in any self-driving system.

The second paper is, honestly, more fun. A team has gotten spiking neural networks running on a five dollar ESP32 microcontroller to control a butterfly-inspired flapping-wing robot weighing less than 30 grams. The paper describes a hierarchical control framework with two lightweight SNNs: one for state estimation from raw sensors, another for control through a central pattern generator that modulates wing actuation. They trained it via imitation learning and achieved stable pitch and heading angle tracking during actual untethered flight.

Let me repeat that: a five dollar chip, two spiking neural networks, a robot that weighs about as much as a AA battery, flying autonomously. The researchers report a 36% latency reduction (1059 microseconds down to 680) and 18% power reduction (0.033 watts to 0.027 watts) compared to a conventional artificial neural network baseline. These aren't the massive gains that neuromorphic evangelists have promised for decades, but they're real gains on real hardware that you can actually buy.

The team claims this is the first demonstration of fully onboard neuromorphic control for autonomous flight of a flapping-wing micro aerial vehicle. I haven't verified that claim exhaustively, but I've been following this space for a while and I can't think of a counterexample, so I'll take their word for it.

Here's what I think is actually happening. For years, neuromorphic computing was a solution looking for a problem. The hardware was exotic, the software tools were immature, and conventional deep learning kept getting faster and cheaper anyway. Why bother with spikes when NVIDIA keeps shipping better GPUs?

But we've hit a wall on the edge. Not everywhere, but in specific domains where power budgets are measured in milliwatts and latency budgets are measured in microseconds. Drones. Wearables. Implantables. Tiny robots. Sensor nodes. The kind of applications where you can't just throw more compute at the problem because there's nowhere to put the compute and nothing to power it with.

I've seen this movie before, actually. It's the same pattern as the early days of mobile computing, when everyone said smartphones would never match desktop performance (true, at the time) and therefore smartphones were toys (wrong, spectacularly wrong). The question was never whether mobile chips would match desktops. The question was whether they'd be good enough for the tasks that mattered on mobile. Turns out they were.

Spiking neural networks might be in that same position now. They don't need to beat transformers on language modeling or diffusion models on image generation. They need to be good enough for specific edge tasks while being dramatically more efficient. A 3.33x energy reduction isn't going to matter in a data center. It might matter quite a lot on a drone that needs to fly for an extra fifteen minutes.

There are caveats, because there are always caveats. Neither of these papers demonstrates anything on actual neuromorphic hardware like Intel's Loihi 2 or the various academic chips floating around. The LiDAR paper simulates energy consumption; the flapping-wing paper runs on conventional silicon (the ESP32) using a software SNN implementation. We don't know yet whether the theoretical efficiency gains translate to real-world neuromorphic deployment, and historically, that translation has been, well, bumpy.

The training story is also still messy. Both papers use surrogate gradient methods, which is basically a hack to make backpropagation work with non-differentiable spike functions. It works, clearly, but it's not elegant and it's not clear whether it scales to much larger networks. The imitation learning approach in the flapping-wing paper sidesteps some of these issues but introduces its own limitations.

And I should note that I only found these two papers because they happened to hit arXiv the same week. This isn't a comprehensive survey of the field. There might be other work that contradicts these results, or extends them in ways I'm not aware of. The neuromorphic literature is fragmented across robotics, neuroscience, and computer engineering venues, and nobody reads all of it.

So what do I actually think? I think we're past the point where neuromorphic computing can be dismissed as academic curiosity. The results are getting too concrete, the applications too specific, the efficiency gains too measurable. But I also think we're nowhere near the point where SNNs replace conventional neural networks for most tasks. This is a niche technology finding its niches, not a paradigm shift.

The interesting question is how big those niches are. If every delivery drone, every agricultural robot, every wearable health monitor, every smart sensor eventually runs some form of neuromorphic perception or control, that's a big market even if it's not the whole market. The kids building these systems today (and yes, I know researchers in their thirties aren't actually kids, but what do I know) are betting that the niches add up to something significant.

I'm not ready to say they're right. But I'm also not ready to say they're wrong anymore, and for someone who's been skeptical of neuromorphic hype since before most of these researchers started grad school, that's actually a pretty big concession.

If you want to argue about this, my email's on the about page. I still prefer it to Twitter.

Aisha Patel · 3 hours ago · 7 min