The Neon Perceptron is a modern, physical interpretation of Rosenblatt’s perceptron in which every wire in the network is a flexible LED that lights up with its own activation. It turns a neural network from an invisible cascade of arithmetic into a glowing object you can stand in front of and watch think.
A neural network you can watch think#
The network has three layers: four inputs you set by drawing on a 2×2 grid, a hidden layer of two neurons, and three outputs. Draw a pattern — a diagonal, a row, a column — and the activations propagate outward along the LED wires, lighting the network up connection by connection until one of the three outputs wins. The colour and intensity of each wire expose exactly what is happening inside, the internal state that, in a digital network, stays buried in floating-point arrays.
The team knows perfectly well that a neural network is not a brain — Kieran Browne, a former PhD student in the School, wrote a thesis arguing exactly that. The name stuck regardless, and the glowing wires really do look like worms.
The Neon Perceptron is a Cybernetic Studio project. The software is by Ben Swift — it runs on a Raspberry Pi using Nerves and Elixir — and the custom PCBs and hardware are designed by Brendan Traw, with the two working on the overall form together.
What a single layer can’t do#
The design holds Rosenblatt’s 1950s perceptron at one end and a pointed piece of AI history at the other. In 1969 Minsky and Papert showed that a single-layer perceptron cannot compute XOR — it manages AND and OR, but not the function that fires only when its inputs disagree — and the field cooled for a decade until non-linear activations and hidden layers made the problem go away. The Neon Perceptron wears that lineage openly. Its task — sorting a 2×2 pattern into a diagonal, a row or a column — is exactly the kind a single-layer perceptron cannot separate, and exactly the kind that two hidden neurons, lighting up in between, can.
A web-based digital twin, rendered in Three.js, mirrors the physical design — draw on the input grid and watch the activations flow without having to stand in the room.
Code and design files: github.com/ANUcybernetics/neon-perceptron.
