The Perceptron Apparatus is a 1.2-metre-diameter wooden instrument with seven concentric rings of sliders. Each ring represents a layer in a simple neural network. You physically slide the rings to perform matrix multiplication and ReLU activation — turning what’s usually an invisible cascade of floating-point arithmetic into something you can see and touch.
Where does the intelligence live?#
Digital computers hide computation behind layers of abstraction. The apparatus does the opposite. It makes you do the work — sliding rings, reading off values, carrying numbers forward through the network. You become the “human computer” in the neural network calculation. And rarely on your own: the apparatus is large enough that a single classification becomes a few-handed affair, one person reading off weights while another slides the rings.
This matters because it forces a strange question. If a neural network is intelligent, where does that intelligence reside? In the weights? In the architecture? In the person sliding the rings? The apparatus sits somewhere between a séance and a slide rule — and that ambiguity is the point. AI is built on an ontology that separates mind from material reality. Making the computation physical and embodied exposes how odd that assumption really is.
How it works#
The seven concentric rings map directly onto a feedforward network: input, weights, hidden layer, more weights, and output. One ring is a logarithmic scale — essentially a circular slide rule — that handles the multiplications.
For MNIST digit recognition, you draw a digit on a worksheet, look up the corresponding weights on a reference poster, slide the rings, and read off the answer. The same apparatus also handles poker hand classification, encoding five-card hands as input features. Two very different domains, one shared physical form.
Every design and fabrication file is openly available in the project repository, so the whole apparatus can be remade from scratch. The trained weights are printed onto A3 reference posters and worksheets you read from as you operate it. Fabrication support came from Sam Shellard at UC Workshop7.
Is it an abacus? Is it an ouija board? No — it’s a perceptron apparatus.
In the lineage of calculating machines#
Not an abacus, then, but a cousin of one. People have built machines to compute for thousands of years: the Antikythera mechanism that tracked the heavens with bronze gears, the abacus, the slide rule, Babbage’s Difference Engine. In each of them the calculation sits right there in the movement of the parts. The Perceptron Apparatus brings a neural network, the engine behind modern AI, into that same visible, mechanical company.
What it gives up is speed. Worked by hand, the apparatus manages something like a tenth of a calculation per second: one multiply-and-add every ten seconds or so. NVIDIA’s current data-centre AI chip, the Blackwell B200, is rated at around twenty petaFLOPS, which is twenty quadrillion such operations every second. What that chip finishes in one second would take the apparatus, worked without pause, something like six billion years, longer than the Earth has existed. The apparatus spends all that speed on the one thing the chip cannot give you: a computation slow enough, and large enough, to watch.
A Cybernetic Studio project by Ben Swift (2024). Code and fabrication files: github.com/ANUcybernetics/perceptron-apparatus.
