Learning quantum state properties with quantum and classical neural networks

Brunello Tirozzi

Bhawani Shankar Leelar, E. S. Shivaleela, T. Srinivas

Artificial Neural Networks (ANNs) are computational models used in Machine Learning (ML), based on a large collection of connected simple units called artificial neurons, loosely analogous to axons in a biological brain. Most of the Neural Networks (NNs) work in classical architecture and a few hardware implementations with optical components have been proposed, which facilitates real-time parallel processing of massive data sets. While these Optical Neural Networks (ONNs) are faster and comparable in efficiency with the classical NNs, they lack in scalability and flexibility. We have proposed a scalable and reconfigurable photonic lattice based NN, similar to feedforward neural network, where photonic lattice represents the layer of the network. The superposition property of quantum computations allows photons to process massive data sets faster than electronic circuits. We explicitly tried to exploit quantum coherence for learning. Hardware implementations put restrictions on the weight space of photon as negative weights are not possible, therefore we have modified the Backpropagation algorithm for optimizing the weights of the network to manipulate the quantum states. The proposed architecture is theoretically scalable as any number of photonic lattice can be included and provides a platform to get benefit of photonic lattices with different properties. It can be used in understanding the learning behaviour of NNs in a constrained environment.

Jincheng Zhang

This paper introduces a novel neural network architecture leveraging the principles of quantum entanglement to enhance parallel computation and learning efficiency. Traditional neural networks are fundamentally limited by the speed of light and the sequential processing of information. We propose a multi-layer parallel neural network where neurons are interconnected through entangled qubits. The synaptic weights are encoded within the entangled states, and the activation function utilizes the superposition property of entangled states for parallel processing. A learning algorithm based on the interference effects within entangled states is developed to adjust the synaptic weights. This approach theoretically enables exponential increases in parallel computation capability and faster learning rates compared to conventional neural networks. The core claim of this work is that quantum entanglement provides a mechanism for instantaneous information transfer between neurons, surpassing the limitations of classical physics and significantly boosting neural network performance.

Huikang Huang, Haozhen Situ

Sanjay Gupta, Royce K. P. Zia

Jincheng Zhang

Quantum Neural Networks (QNNs) represent a promising paradigm shift in machine learning, leveraging the unique properties of quantum mechanics to potentially surpass classical neural networks. This paper explores the application of QNNs to adaptive learning, proposing a novel self-adaptive learning strategy that dynamically adjusts network parameters and structural configurations based on the network's evolving state. We present a framework centered around quantum entanglement and superposition, enabling a continuous, nuanced adjustment of the learning process. The core mechanism focuses on leveraging the inherent quantum properties to optimize learning efficiency. We demonstrate the efficacy of this approach through simulations and preliminary experiments, illustrating a significant improvement in learning speed and robustness compared to traditional approaches. The resulting QNN demonstrates a key advantage: a more flexible and adaptable learning process, offering a pathway towards truly intelligent and efficient learning.