Quantum Entanglement in Deep Learning Architectures.

Area of Science:

• Quantum Information Science
• Machine Learning
• Computational Physics

Background:

• Deep learning has achieved significant advancements across various fields.
• Machine learning for wave function representation traditionally uses restricted Boltzmann machines (RBMs) and fully connected networks.

Purpose of the Study:

• To investigate the efficacy of contemporary deep learning architectures for representing highly entangled quantum systems.
• To compare the entanglement capacity of advanced neural networks against traditional machine learning models.

Main Methods:

• Utilized deep convolutional and recurrent neural networks for quantum wave function representation.
• Constructed tensor network equivalents of these deep learning architectures.
• Analyzed information reuse within network operations.

Main Results:

• Deep convolutional and recurrent networks efficiently represent highly entangled quantum systems.
• These architectures exhibit inherent information reuse, distinguishing them from standard tensor network methods.
• The networks demonstrate superior entanglement capacity, supporting volume-law entanglement scaling.

Conclusions:

• Contemporary deep learning architectures offer a polynomially more efficient approach to wave function representation compared to RBMs.
• The findings motivate the adoption of state-of-the-art machine learning models for quantum wave function representations.
• This work quantifies the entanglement capacity of leading deep learning architectures.