Perceptron apparatus: inference walkthrough

19 Mar 26

aistudio

The perceptron apparatus is a 1.2-metre diameter circular wooden table that classifies handwritten digits using concentric rings of sliders and a logarithmic slide rule. It’s a piece of speculative design from the School of Cybernetics at ANU, imagined as an artefact from a world where the knowledge to build digital computers has been lost, but the mathematics of neural networks survived. So people built this instead.

The apparatus implements a real trained multilayer perceptron—a 36→6→10 MLP that takes a 6×6 pixel image of a digit, runs a forward pass through its weights and activations, and produces a prediction. Every multiply is a slide rule operation. Every accumulation is a slider moving along a track. The whole inference process is physically legible, each step visible to anyone standing around the table.

The physical table has been built—fabricated by Sam Shellard at UC’s Workshop7, with all the laser cutting and CNC routing files available in the GitHub repo. This post walks through the inference process with an interactive digital twin of the apparatus. If you’ve seen the neon perceptron posts, this is its older, more analogue sibling.

The architecture

The apparatus has five concentric rings, each corresponding to a layer of the network:

• Ring A (outermost): 36 radial sliders for the input pixels—one per cell in the 6×6 grid
• Ring B: weight sliders for the input→hidden connections (36×6 = 216 weights)
• Ring C: 6 radial sliders for the hidden neuron activations
• Ring D: weight sliders for the hidden→output connections (6×10 = 60 weights)
• Ring E (innermost): 10 radial sliders for the output—one per digit class

Around the outside sit two additional rings: a logarithmic rule ring for performing multiplication, and a ReLU reference ring for reading off the activation function. To multiply two numbers you rotate the log ring to align the two values, then read the product off the outer scale—the same principle as a slide rule, wrapped into a circle.

The forward pass is:

hidden = relu(input × B)
output = hidden × D
prediction = argmax(output)

No bias terms, no softmax on the output. The hidden layer uses ReLU—negative values snap to zero, which on the apparatus just means the slider stays at the bottom of its track.

Try it

Pick a digit from the thumbnails below and hit “Step through” to watch the forward pass animate on the apparatus. Each step computes one hidden or output neuron—accumulating the weighted inputs, applying ReLU where appropriate, and sliding the result into place. “Instant” skips the animation and shows the final state.

The weights are real—trained on MNIST via Axon in Elixir, then exported to the digital twin. The five sample digits (0, 1, 4, 6, 7) are ones the model classifies correctly. With only 276 parameters and aggressive downsampling from 28×28 to 6×6 pixels, it’s not going to win any accuracy benchmarks—but that was never the point. The apparatus exists to make the forward pass visible.

Related posts

The apparatus is one of several physical neural network projects at the School of Cybernetics. The others use a different aesthetic—LEDs and PCBs rather than wood and brass—but explore the same idea of making computation tangible:

• Interactive neon perceptron visualisation—a 5×5 input, 9-hidden, 10-output network rendered in Three.js
• Mini neon perceptron: XOR edition—a tiny 2×2 input version for testing the physical build

@online{swift2026perceptronApparatusInferenceWalkthrough,
  author = {Ben Swift},
  title = {Perceptron apparatus: inference walkthrough},
  url = {https://benswift.me/blog/2026/03/19/perceptron-apparatus-inference-walkthrough/},
  year = {2026},
  month = {03},
  note = {AT-URI: at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.document/2026-03-19-perceptron-apparatus-inference-walkthrough},
}