#lieprune

Lieprune compresses quantum neural networks by 10x without affecting machine learning performance.

Quantum neural network is an intriguing machine learning method for the near future. Due to the large number of parameters needed to function, computational hurdles limit these networks' potential. Scalability constraints, barren plateaus, and technology limitations make near-term quantum machine learning difficult.

In a major effort to address these challenges, Jiangsu University of Science and Technology scientists Haijian Shao, Bowen Yang, and Wei Liu, along with University of Nevada, Las Vegas professor Yingtao Jiang, and his colleagues, presented LiePrune. A mathematically-based one-shot structured pruning framework, LiePrune is ideal for parameterised quantum circuits and quantum neural networks. This project aims to simplify these complex networks for scalable, practical quantum machine learning.

Redundancy Detection Principles

The innovative combination of quantum geometry and Lie group theory distinguishes LiePrune. This innovative, mathematically based methodology allows the framework to locate and delete superfluous parameters consistently, resulting in aggressive network compression.

The core mechanism includes each gate. This representation includes a quantum geometric feature space, Lie group, and corresponding Lie algebra dual space. This novel dual representation method detects redundancy efficiently, allowing quantum circuits to be aggressively compressed without compromising their critical functionality. The Lie group structure of quantum circuits allows LiePrune to reduce parameters significantly.

The research team proved that LiePrune offers verified assurances and great compression. These promises about functional approximation, redundancy detection, and computer efficiency advance scalable and effective quantum machine learning.

Classification shows aggressive compression The researchers found that LiePrune can compress models for several classification tasks without losing accuracy. Quantum classification tasks were tested on the prominent MNIST and FashionMNIST datasets. The study found that LiePrune can compress models by 8-10 times, with some instances exceeding 10×.

On the MNIST 4-vs-9 dataset, the team reduced parameters from 288 to 36. Despite this substantial parameter decrease, the network preserved 95.9% of its initial accuracy after a fast fine-tuning.

The Fashion Sandal-vs-Boot dataset yielded similar promising results with LiePrune. The framework cut classification criteria from 360 to 36. The model was 74.0% accurate after correction. These results prove that LiePrune compresses quantum models well for classification.

Quantum Chemistry Simulation Sensitivity

The research team used LiePrune to solve the quantum chemistry assignment LiH Variational Quantum Eigensolver (VQE) to broaden their study. Using a 12-qubit, 12-layer ansatz, LiePrune compressed this domain 12-fold, reducing the number of parameters from 432 to 36.

In contrast to benchmark classification tasks, quantum chemistry findings were more sensitive to strong pruning. Due to the severe 12-fold compression, the estimated energy deviation first declined.

Even after further fine-tuning restored the ground state energy, a 3.23 Ha gap remained. Further analysis found that extremely minor compression levels caused energy aberrations that could be fully recovered with fine-tuning. The extraordinarily intense compression caused chemically structured Hamiltonians to be inaccurate.

Potential Scalability and Future Work

LiePrune accelerates the development of practical quantum neural networks and parameterised quantum circuits. Trimming these circuits solves the system's scalability problem caused by many parameters and high processing needs. This breakthrough in applying quantum computing principles to answer complex operations tenfold faster than traditional computers is the ability to lower parameters by eight to twelve times, often with modest or increasing performance.

Even though classification tasks were successful, chemically structured Hamiltonians were more sensitive, indicating that more improvements are needed. The results indicate that this field requires unique techniques to maintain precision. Future research should focus on integrating enhancements like chemistry-aware restrictions to maximise LiePrune's benefits in large simulations like VQE.