| （英） |
Reservoir-computing systems consist of an input layer, the reservoir, and the output layer. As opposed to the convolutional neural network which needs all the weights to be optimized for feature extraction, network connections of the reservoir are randomized and remain unchanged throughout computation. Only the output layer is trained, typically with a linear regression, which greatly minimizes computational resources. For complex dynamical responses, the reservoir implements a high-dimensional nonlinear mapping to the output layer from the input layer. We numerically analyze a photon-based quantum scrambler consisting of multi-stage beam splitters and phase shifters as a candidate of the bosonic quantum reservoir. Complex dynamical responses are created by non-classical nature of photons (multi-photon interference and entanglement) and nonlinear effects of photon coincidence measurement. Furthermore, the bosonic nature of photons allows expansion of the calculation (Hilbert) space by simply adding photons to the reservoir without changing its physical configuration. However, entanglement is easily broken by optical loss; thereby, we usually observe both classical and non-classical states as a mixture. Our numerical results suggest that such a mixed state can help fully use the calculation space of a photon-based bosonic quantum reservoir. |