Lieprune Achieves Over 10x Compression of Quantum Neural Networks with Negligible Performance Loss for Machine Learning Tasks
One extremely interesting approach for near-term machine learning applications is the use of quantum neural network (QNN). However, significant computational obstacles now limit these networks’ full potential, mainly because to the enormous number of parameters needed for functioning. In addition to problems like barren plateaus and limitations in existing technology, these scalability constraints provide major challenges for near-term quantum machine learning.
A group of scientists from Jiangsu University of Science and Technology, Haijian Shao, Bowen Yang, and Wei Liu, along with Yingtao Jiang from the University of Nevada, Las Vegas, and their associates, have presented LiePrune as a novel solution in a significant attempt to overcome these obstacles. LiePrune is a one-shot structured pruning framework with a mathematical foundation that is well suited for parameterized quantum circuits and quantum neural networks. The aim of this effort is to significantly reduce these intricate networks in order to achieve scalable, useful quantum machine learning.
You can also read Quantum Nexus Powers California’s Quantum Research Regime
A Principled Approach to Redundancy Detection
LiePrune’s framework sets itself apart by merging concepts from quantum geometry and Lie group theory in a novel way. Aggressive network compression results from the framework’s ability to discover and then delete unnecessary parameters in a principled manner according to this advanced, mathematically based methodology.
Each gate is jointly represented as part of the fundamental mechanism. This representation covers a quantum geometric feature space, a Lie group, and the dual space of a Lie algebra that corresponds to it. This innovative dual representation method enables extremely effective redundancy detection, allowing quantum circuits to be aggressively compressed without sacrificing the circuit’s essential functioning. LiePrune can achieve significant decrease in parameters by using the underlying Lie group structure of quantum circuits.
The research team showed that LiePrune provides proven assurances in addition to attaining high compression. A significant step towards scalable and useful quantum machine learning is provided by these assurances, which are notably related to functional approximation, redundancy detection, and overall computing efficiency.
You can also read CMTS Cryogenic Muon Tagging System for Quantum Processors
Demonstrating Aggressive Compression in Classification The researchers’ experiments demonstrate LiePrune’s capacity to compress models on a variety of classification tasks with little loss of accuracy. The popular MNIST and FashionMNIST datasets were used to test the framework on quantum classification tasks. The findings, LiePrune can compress models by a factor of 8 to 10, and in certain situations, the total obtained compression exceeds 10×.
The team was able to successfully cut the number of parameters from 288 to just 36 on the MNIST 4-vs-9 dataset. Importantly, after a quick fine-tuning procedure, the network retained 95.9% of its initial accuracy despite this drastic parameter reduction.
Comparable encouraging outcomes were noted when LiePrune was used on the Fashion Sandal-vs-Boot dataset. The framework reduced the criteria in this classification problem from 360 to 36. The model attained 74.0% accuracy after adjustment. These results unequivocally show that LiePrune is very good at compressing quantum models for classification applications.
Sensitivity in Quantum Chemistry Simulations
The research team expanded the scope of their study by using LiePrune to solve the LiH Variational Quantum Eigensolver (VQE) issue, a quantum chemistry assignment. By employing a 12-qubit, 12-layer ansatz, LiePrune accomplished a remarkable 12-fold compression in this domain, significantly lowering the number of parameters from 432 to 36.
In contrast to the benchmark classification tasks, the quantum chemistry results showed a higher sensitivity to strong pruning. The computed energy deviation first deteriorated significantly due to the extreme 12-fold compression.
There was still a discernible 3.23 Ha gap even after further fine-tuning largely restored the ground state energy subsequent investigation revealed that very slight compression levels resulted in energy aberrations that could be completely recovered with fine-tuning. However, the chemically structured Hamiltonians exhibited significant inaccuracies due to the extremely strong compression.
You can also read Infleqtion’s Contextual Machine Learning for the U.S. Army
Implications for Scalability and Future Work
A major step forward in the creation of useful quantum neural networks and parameterized quantum circuits is represented by LiePrune. The system overcomes a significant scalability constraint brought on by an abundance of parameters and high processing needs by effectively trimming these circuits. This significant advancement in using quantum computing principles to answer complicated computations tenfold faster than conventional computers is the capacity to lower parameters by factors of more than eight to twelve times, frequently with modest or even increased performance.
The greater sensitivity seen while working with chemically structured Hamiltonians indicates that more improvements are required, even if classification tasks were completed with overwhelming success. The results suggest that specific tactics are needed to completely maintain accuracy in this field. To guarantee that the full advantages of LiePrune can be realized throughout large simulations like VQE, the researchers recommend that future work concentrate on integrating enhancements like chemistry-aware limitations.
The LiePrune Quantum Geometric Dual Representation for One-Shot Structured Pruning of Quantum Neural Networks, demonstrates the rapid advancement of quantum research and establishes LiePrune as an essential resource for individuals seeking to unleash the potential of quantum technology to address unsolvable issues in a variety of industries.
You can also read Microsoft With Algorithmiq To Develop Quantum Chemistry
