| （英） |
Extreme learning machine (ELM) is a feedforward neural network with a single or multiple hidden layers. Unlike the convolutional neural network, node connections of the ELM are randomized and remain unchanged throughout computation. Only, the output layer is trained with a simple calculation (a linear regression, for example), greatly minimizing computational resources. For feature extraction, the ELM should implement a high-dimensional nonlinear projection to the output layer from the input layer. We numerically analyze a multi-stage multi-photon phase-randomized interferometer as a photon-based quantum ELM. Feature vectors are created from multi-photon coincidence measurement. Due to the bosonic nature of photons, vector space expansion is achieved by simply adding photons to the ELM. Multi-photon interference and coincidence measurement are the origin of complex dynamical nonlinearity. The bosonic nature of photons contributes to its high-dimensionality. Vector spaces created by photons with and without entanglement are compared to observe their non-classicality in nonlinear projection. A clear trade-off relationship was found between non-classicality and high-dimensionality; however, we propose a simple classical nonlinear operation to keep their balance. |