An AI model whose parameters and state are held in qubits, and which runs on the quantum processor itself.
Training a conventional model means pushing a global error signal backwards through the whole network. That is natural on a GPU and awkward on a processor where reading the state is what destroys it.
We started somewhere else.
The model is built from small cells, each holding a few qubits, and each cell adjusts itself from what it can see around it. Learning is physical relaxation rather than a training pass.
The cells sit in a two-dimensional fabric, so the model grows by adding cells rather than layers. Capacity and training cost stop being tied together.
Quantum information theory already defines what to measure about a quantum state, so the inside of the model is open to inspection as it runs.
It is also what lets a model certify its own answer, and refuse. Our paper
On silicon
Hardware agnostic. Superconducting and trapped ion, ten qubits to over a hundred, learned on one machine and executed on another manufacturer’s.
Technical material and experimental results are available on request.