· snn · brain · neuromorphic hardware
brains don’t run on batteries the size of cities
spiking neural networks and neuromorphic chips try to copy the brain’s quiet, spike-based way of thinking, so ai could one day run on a few watts instead of a power plant.
here’s something that bugs me. the human brain runs on about 20 watts, roughly what a dim light bulb uses, and it does all of it: recognising faces, holding a conversation, remembering where you left your keys. meanwhile, training a big ai model can eat as much electricity as a small town. somewhere along the way we built machines that think in a completely different way than we do, and we’re paying for it.
spiking neural networks are an attempt to get back to how the brain actually works. normal neural networks pass smooth numbers between layers all the time, every neuron talking constantly. real neurons don’t do that. they sit quietly, build up charge, and only fire a short spike when something matters. most of the time they’re silent. an snn copies that idea. information travels as spikes, and if nothing is happening, nothing gets computed.
the learning can be brain-like too. a rule called stdp strengthens a connection when one neuron fires just before another, and weakens it otherwise. no labels and no giant error signal flowing backwards through the whole network. each connection just learns from what happens around it. i’ve been working with this in my own research, and it’s strange and kind of beautiful to watch filters slowly organise themselves into edge detectors without anyone telling them to.
the catch is that ordinary computer chips are terrible at this. a gpu is built to crunch huge blocks of numbers at once, so it ends up simulating every silent moment of every neuron, which wastes the whole point. that’s where neuromorphic hardware comes in. chips like intel’s loihi, ibm’s truenorth and the spinnaker machine in manchester are built around spikes. memory and computing sit together, like they do in the brain, and power is only used when a spike actually happens.
it’s not magic yet. snns still lose to standard networks on accuracy, and the tools are young. but if we want ai that can live on a phone, a sensor or a tiny robot without needing a power plant behind it, learning to think in spikes might be the way there.