Fig. 4: Quantum reservoir computing.
a, Classical neural approach to a classification problem. The input information (for example, the pixels of an image) is fed to the first layer. The network can be trained to ‘switch on’ the neuron in the last layer corresponding to the correct class. b, Reservoir computing. The input information is mapped to a nonlinear, high-dimensional space, whose output is interpreted by a simple linear readout network. Only the readout network is trained, which requires minimal resources. c, A quantum reservoir computer based on quantum memristors. The input information is encoded on the quantum states of three photons in nine optical modes. A fixed matrix of beamsplitters with random reflectivity distributes the information across all the optical modes, which are fed to three quantum memristors, whose outputs are scrambled again before reaching the photon counters. The re-injection of photons is assumed if the measurement takes place at the feedback port of a memristor. The output pattern is interpreted by a trainable linear readout function.
