Quantum generalisation of feedforward neural networks

摘要: We propose a quantum generalisation of classical neural network. The neurons are firstly rendered reversible by adding ancillary bits. Then they generalised to being reversible, i.e.\ unitary. (The networks we generalise called feedforward, and have step-function activation functions.) network can be trained efficiently using gradient descent on cost function perform generalisations tasks. demonstrate numerically that it can: (i) compress states onto minimal number qubits, creating autoencoder, (ii) discover communication protocols such as teleportation. Our general recipe is theoretical implementation-independent. neuron module naturally implemented photonically.