|
Abstract
|
Linear optics neural networks or optical neural networks offer potential advantages over traditional electronic
neural networks in terms of speed, energy efficiency, scalability, and improved parallelism, particularly for highbandwidth
applications. The use of photonics allows for more compact and integrated neural network designs,
potentially enabling the development of larger and more complex networks. A linear optics network is developed
to implement a quantum classifier. Indeed, the designed network is a quantum circuit consisting of some
Gaussian gates such as displacement, noiseless linear amplification (NLA), squeezer and Green machine. At first,
the classical inputs are encoded with the help of position-displacement operator to prepare single-mode coherent
states. Then, the amplitudes of the coherent states are amplified by passing through NLA elements followed by
squeezer gates that may transform classical coherent states into nonclassical ones. Finally, the transformed coherent
states are fed into the Green machine which provides entangled states as the outcome of the network. As
a primary goal of this work, the network generates a multi-mode entangled state by applying the displacement
operator on the vacuum state encoded classical data. Besides, it is shown that the output state of the circuit may
possess squeezing characteristics as another nonclassical feature. In the continuation, as a practical application,
the network is implemented to perform some pattern recognition tasks. At first, the Bayes theorem is employed
to define discriminant functions to perform a general classification task, then the outcome distribution of the
network is utilized to classify some corrupted LEDs that display English letters. Finally, we show that the outcome
of the circuit may be manipulated to embed classical neural networks into a continuous-variable variational
quantum circuit (VQC). The network is trained via the logistic regression algorithm with the MNIST database
|