A Unitary Weights Based One-Iteration Quantum Perceptron Algorithm for Non-Ideal Training Sets

Wenjie Liu, Peipei Gao, Yuxiang Wang, Wenbin Yu, Maojun Zhang

TL;DR

This paper introduces a novel quantum perceptron algorithm that can learn from "non-ideal" training sets (incomplete or over-complete) in a single iteration. By employing Singular Value Decomposition (SVD) to ensure unitary weights, the algorithm accurately implements arbitrary quantum gates, offering a significant advantage over previous methods.

What Problem Does This Paper Address?

The development of quantum neural networks (QNNs) is an active research area, aiming to leverage quantum computation's power for machine learning tasks. A key challenge in building effective QNNs, particularly perceptrons (the simplest neural network unit), is handling imperfect or incomplete training data. Traditional perceptron algorithms often require iterative learning and can struggle with non-ideal datasets. Furthermore, many proposed quantum perceptron models either don't maintain unitary evolution (a fundamental requirement in quantum mechanics) or are limited in the types of quantum gates they can implement. This paper tackles these issues by proposing a quantum perceptron that is both robust to non-ideal training sets and capable of implementing any quantum gate in a single, unitary step.

Methods

The core innovation of this work lies in its approach to weight matrix construction and its use of Singular Value Decomposition (SVD). The algorithm first calculates the "total weight" from the training data. Crucially, instead of directly using this weight matrix, which may not be unitary, they decompose it using SVD into three matrices: $M = U \Sigma V^*$. The non-unitary weight matrix $\hat{w}$ is then transformed into a unitary form by replacing the diagonal matrix of singular values ($\Sigma$) with a unitary matrix ($\Sigma_{new}$) composed of ones on the diagonal. This ensures that the resulting quantum perceptron operation is unitary. The final output is then computed by applying this unitary transformation to the input quantum state. This entire process is designed to be completed in a single iteration.

Datasets

The paper does not explicitly mention or introduce new datasets. Instead, it uses synthetic training sets constructed to represent various quantum gates (Hadamard, S, T, CNOT, Toffoli, Fredkin) and different training set scenarios: complete, less-complete, and over-complete. These examples serve as the "datasets" for validating the algorithm's performance.

Key Insights and Ideas

  • One-Iteration Learning for Quantum Perceptrons: The ability to achieve accurate learning and gate implementation in a single iteration is a significant leap, potentially speeding up quantum machine learning training considerably.
  • Robustness to Non-Ideal Training Data: The algorithm's success with less-complete and over-complete training sets addresses a practical limitation of many existing quantum machine learning models, making them more applicable to real-world scenarios where perfect data is rare.
  • Unitary Weight Matrix via SVD: The clever use of SVD to transform a potentially non-unitary weight matrix into a unitary one is the key to ensuring the algorithm adheres to quantum mechanical principles while still learning from data. This is a novel approach to maintaining unitarity in a learning context.
  • Universal Quantum Gate Implementation: Demonstrating the algorithm's ability to accurately implement a range of universal quantum gates, including complex ones like Toffoli and Fredkin, highlights its potential for general quantum computation tasks.

Key Results

The paper demonstrates that their proposed quantum perceptron algorithm can accurately implement various quantum gates (H, S, T, CNOT, Toffoli, Fredkin) within a single iteration. This accuracy is maintained even when using less-complete or over-complete training sets, a capability not shared by the reference algorithms (Altaisky's and Seow et al.'s). The performance evaluation shows that their algorithm offers higher availability and wider applicability compared to the other methods, achieving accurate results in one iteration where others might require more or fail entirely on non-ideal data.

Limitations and Future Work

The authors acknowledge that their current algorithm is designed for single-layer perceptrons. They suggest that future research should focus on extending these techniques to multi-layer perceptrons, which would likely require more sophisticated algorithms akin to backpropagation. Additionally, they propose exploring the application of their quantum perceptron to construct larger-scale quantum neural networks by combining multiple perceptron units.

Summary updated

Abstract

In order to solve the problem of non-ideal training sets (i.e., the less-complete or over-complete sets) and implement one-iteration learning, a novel efficient quantum perceptron algorithm based on unitary weights is proposed, where the singular value decomposition of the total weight matrix from the training set is calculated to make the weight matrix to be unitary. The example validation of quantum gates {H, S, T, CNOT, Toffoli, Fredkin} shows that our algorithm can accurately implement arbitrary quantum gates within one iteration. The performance comparison between our algorithm and other quantum perceptron algorithms demonstrates the advantages of our algorithm in terms of applicability, accuracy, and availability. For further validating the applicability of our algorithm, a quantum composite gate which consists of several basic quantum gates is also illustrated.
BibTeX
@inproceedings{liu2019unitary,
  author = {Wenjie Liu and Peipei Gao and Yuxiang Wang and Wenbin Yu and Maojun Zhang},
  title = {{A Unitary Weights Based One-Iteration Quantum Perceptron Algorithm for Non-Ideal Training Sets}},
  booktitle = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2896316},
}
Figure