Reviewed by Lexie CornerNov 11 2024
The field of quantum machine learning, which aims to leverage quantum computing to improve upon classical machine learning, has made notable progress recently. A study published in Intelligent Computing by researchers from the University of Melbourne and CSIRO shows that the complexity of quantum circuits for data encoding in quantum machine learning can be reduced without sacrificing robustness or accuracy.
This diagram illustrates how the team reduces quantum circuit complexity in machine learning using three encoding methods—variational, genetic, and matrix product state algorithms. All methods significantly reduce circuit depth while preserving accuracy, as shown by the histogram of reduced CNOT gate counts. Image Credit: Maxwell T. West et al.
The team's findings, confirmed through simulations and experiments on IBM quantum devices, show that their novel encoding techniques achieved comparable classification accuracy while reducing circuit depth by an average of 100 compared to traditional methods.
These results offer a promising direction for the practical implementation of quantum machine learning on current quantum devices. Looking ahead, the team aims to explore further optimizations in quantum state encoding and machine learning architecture, as well as scale their models to handle larger and more complex datasets.
Encoding classical data into quantum states has been a significant challenge in quantum machine learning, as it requires deeply entangled circuits.
To address this, the team developed three encoding methods that approximate the quantum state of the data with much shallower circuits. These methods—matrix product state, genetic, and variational algorithms—preserved classification accuracy on the MNIST image dataset and two others while increasing resilience to adversarial data manipulation.
Effective state preparation was achieved through the unique quantum state encoding methods for classical data:
-
Matrix product state encoding: This method generates sequentially disentangled quantum states using tensor networks. It further reduces complexity by allowing quantum states to be prepared with low entanglement and a limited number of Controlled-NOT (CNOT) gates.
-
Genetic algorithm for state preparation: Inspired by evolutionary processes, this method optimizes state preparation by producing multiple circuit configurations, selecting the most effective one, and reducing the number of CNOT gates to enhance the circuits' noise resistance.
-
Variational coding: This technique uses trainable parameters within a layered circuit structure to enable quantum states to reach the desired accuracy with fewer layers, typically lowering computational costs and minimizing the need for extensive entangling operations.
The study aligns with broader objectives in quantum machine learning, aiming to develop efficient and reliable quantum models for applications such as image recognition, cybersecurity, and complex data analysis.
Reducing circuit depth is crucial for enabling practical quantum machine learning on current devices, which are often constrained by gate fidelity and qubit count. Additionally, the enhanced resilience of these models to adversarial attacks paves the way for secure quantum machine learning applications in industries where tamper resistance is essential.
Journal Reference:
West, M. T. et. al. (2024) Drastic Circuit Depth Reductions with Preserved Adversarial Robustness by Approximate Encoding for Quantum Machine Learning. Intelligent Computing. doi.org/10.34133/icomputing.0100