Quantum-Inspired Stacked Auto-encoder-Based Deep Neural Network Algorithm (Q-DNN)

Abstract

In this paper, a quantum-inspired stacked auto-encoder-based deep neural network (Q-DNN) learning algorithm is proposed. The proposed Q-DNN uses stacked auto-encoder to form a deep neural network. This quantum computing concept has been used to optimize the learning parameters of the algorithm. To form a deep neural network, the selection of learning algorithms and its parameters is an important criterion for achieving better performance. However, learning capability depends upon the proper selection of learning parameters; improper selection may cause local minima or local maxima problem. To avoid improper selection of learning algorithm parameters, the quantum computing concept has been used, which is characterized by the representation of population dynamic, and evaluation function. It uses qubit for probabilistic representation which gives a large subspace for exploration to find an optimal value of the learning parameters. This concept has been utilized in the formation of stacked auto-encoder-based deep neural network by optimizing learning parameters of the learning algorithm. The proposed Q-DNN is trained and tested on various benchmark datasets such as the BUPA Liver Disorder, Ionosphere, PIMA Indians Diabetes, and the MNIST dataset on three different neural network architectures. Firstly, the deep neural network architecture is formed with one hidden layer, then deep neural network formed with two hidden layers, and finally deep neural network formed with three hidden layers. The proposed Q-DNN achieves promising results in terms of classification accuracy, sensitivity, and specificity in comparison with other approaches.