Perceptron theory · another classical body · mechanical
The perceptron in sound
Light's slow cousin. A surface acoustic wave ripples across a piezoelectric chip at a few thousand metres per second — a hundred-thousand times slower than light, which is exactly the point: the same interference that multiplies, but crawling, so a useful delay fits on a fleck of quartz. Tap the wave at intervals, weight each tap by how much electrode overlaps it, and add — that's a weighted sum in moving matter. A matched filter, which is to say a perceptron, made of ripples.
input = a wave packet · weights = interdigital tap overlaps · multiply+sum = tapped interference Σwᵢ·x(t−τᵢ) · fire = a correlation peak
✓ STRONG
The SAW correlator. Tapped-delay-line filters and matched filters are in every phone's RF front-end — decades-mature. The weighted-sum-of-delays is bedrock signal physics.
◐ MIDDLING
Acoustic neuromorphic. Phononic reservoir computing and SAW neural elements are demonstrated — compact and passive, but niche and bandwidth-bound.
◔ FRONTIER
Large acoustic nets. Scaling phonon computing into real learning machines is open — sound is a memory and a filter more than a deep net, so far.
I · A wave, tapped and weighted
A wave packet travels left to right across the substrate. Fixed taps read its height as it passes; each tap carries a weight (the electrode overlap). At any instant the readout is y = Σ wᵢ · x(t − τᵢ) — the wave sampled at staggered delays, scaled, summed. The slowness is the feature: at 3,488 m/s a 1 GHz wave is only 3.5 µm long, so a whole pattern of delays fits in microns.
the surface wave on the piezo substrate · taps (heights = weights) sample it at delays · the bar = Σwᵢx(t−τᵢ)
0.00
weighted sum at the output tap
multiply= tap weight × wave heightsum= the electrodes add chargedelay= the wave's travel time
No clock drives the sum — the wave's own motion stages the delays. A SAW chip is a passive analogue convolver: the math is done by geometry and time-of-flight.
II · A matched filter is a perceptron
Freeze the weights to a stored pattern. Now slide an input wave through: when its shape lines up with the weights, every tap reinforces and the output spikes — the neuron fires. When it doesn't, the taps cancel to near zero. Correlation is the weighted sum, and a matched filter is a threshold neuron tuned to one pattern.
stored pattern:
input sliding past the stored weights · the correlation output peaks at alignment = the fire
listening…
fires at a correlation peak
III · Learns, and cascades for depth
Train the taps and the acoustic neuron learns AND/OR; one filter is one cut, so XOR needs a second SAW stage fed by a nonlinearity — depth, in cascaded chips. The gift this body gives is memory: a slow wave stores a signal in flight, so an acoustic net carries its own short-term recall in the delay itself.
target:
—press train
the gift= the delay is memorythe limit= bandwidth, RF-scalethe role= filter/correlator, not deep net
Honest place in the family: sound is brilliant at one matched filter and at holding a signal still, less so at stacking into a big network. It's a memory and a feature detector cast in ripples — useful, bounded, real.