Continuous variable quantum perceptron

Abstract

We present a model of Continuous Variable Quantum Perceptron (CVQP) whose architecture implements a classical perceptron. The necessary non-linearity is obtained via measuring the output qubit and using the measurement outcome as input to an activation function. The latter is chosen to be the so-called ReLu activation function by virtue of its practical feasibility and the advantages it provides in learning tasks. The encoding of classical data into realistic finitely squeezed states and the use of superposed (entangled) input states for specific binary problems are discussed.

Opportunities in Quantum Reservoir Computing and Extreme Learning Machines

“…We now move to discuss the substrates that are usually modeled as continuous variables, as for instance optical systems, already well‐established versatile and powerful classical RC platforms. [ 64 ] As for machine learning with quantum continuous variables, proposals to realize either basic building blocks [ 81 ] or full scale quantum NN [ 76 ] using variational (i.e., parameterized) quantum circuits have recently been reported. In these proposals, linear transformations and displacements of the input are achieved by Gaussian gates.…”

“…Measurement‐induced nonlinearity is proposed as an alternative to experimentally difficult non‐Gaussian gates in ref. [81], however no advantage over classical perceptrons was observed. Although intended to be trained by adjusting the circuit parameters, these proposals could be adapted to realize continuous‐variable QELM or QRC by starting from a randomly initialized but fixed feedforward or recurrent architecture, respectively, and training only the readout layer.…”

Opportunities in Quantum Reservoir Computing and Extreme Learning Machines

“…We now move to discuss the substrates that are usually modelled as continuous variables. Proposals to realize either basic building blocks [71] or full scale quantum NN [67] using variational (i.e. parameterized) quantum circuits have recently been reported.…”

“…Nonlinearity is introduced with non-Gaussian gates in [67]. Measurement-induced nonlinearity is proposed as an alternative to experimentally difficult non-Gaussian gates in [71], however no advantage over classical perceptrons was observed. Although intended to be trained by adjusting the circuit parameters, these proposals could be adapted to realize continuous variable QELM or QRC by starting from a randomly initialized but fixed feedforward or recurrent architecture, respectively, and training only the readout layer.…”

Pattern capacity of a single quantum perceptron

“…The simplest model of an artificial neuron traces back to the classical Rosenblatt's perceptron [6], which can be seen as the simplest learning algorithm for binary classification. Several possiblities can be considered to implement a perceptron by means of a quantum architecture [7,8,9,10,11,12,13]. In this context it is important to investigate the capability of a particular model of quantum perceptron to achieve quantum advantages with respect to its classical counterparts.…”