[Paper Review] Quantum Gauged Neural Network: U(1) Gauge Theory

Y. Fujita, Tetsuo Matsui|arXiv (Cornell University)|Jun 30, 2002

Neural Networks and Applications4 citations

TL;DR

This paper proposes a quantum extension of the classical Z(2) gauged neural network by replacing discrete spin and coupling variables with continuous U(1) phase variables, forming a U(1) lattice gauge theory analogous to Ginzburg-Landau superconductivity. Using mean-field theory, it identifies Coulomb, Higgs, and confinement phases and finds that quantum fluctuations weaken learning and recall abilities compared to the classical Z(2) model.

ABSTRACT

A quantum model of neural network is introduced and its phase structure is examined. The model is an extension of the classical Z(2) gauged neural network of learning and recalling to a quantum model by replacing the Z(2) variables, $S_i = \pm1$ of neurons and $J_{ij} =\pm1$ of synaptic connections, to the U(1) phase variables, $S_i = \exp(iϕ_i)$ and $J_{ij} = \exp(iθ_{ij}) $. These U(1) variables describe the phase parts of the wave functions (local order parameters) of neurons and synaptic connections. The model takes the form similar to the U(1) Higgs lattice gauge theory, the continuum limit of which is the well known Ginzburg-Landau theory of superconductivity. Its current may describe the flow of electric voltage along axons and chemical materials transfered via synaptic connections. The phase structure of the model at finite temperatures is examined by the mean-field theory, and Coulomb, Higgs and confinement phases are obtained. By comparing with the result of the Z(2) model, the quantum effects is shown to weaken the ability of learning and recalling.

Motivation & Objective

• To explore the differences between classical and quantum neural networks by extending the Z(2) gauged neural network to a U(1) gauge-theoretic framework.
• To investigate how quantum fluctuations, arising from continuous U(1) variables, affect learning and memory recall capabilities.
• To establish a phenomenological quantum model of neural networks that mirrors aspects of lattice gauge theories and superconductivity.
• To examine the emergence of conserved currents in the quantum model and their potential role in modeling signal propagation in neural systems.

Proposed method

• Replace classical Z(2) variables $S_i = \pm 1$ and $J_{ij} = \pm 1$ with U(1) phase variables $S_i = e^{i\varphi_i}$ and $J_{ij} = e^{i\theta_{ij}}$, forming a U(1) lattice gauge theory.
• Construct an energy functional invariant under local U(1) gauge transformations, ensuring local dynamics analogous to gauge field theories.
• Apply mean-field theory to analyze the phase structure of the model at finite temperature, identifying distinct phases such as Coulomb, Higgs, and confinement.
• Define a gauge-invariant current $j_{x\mu}$ via Noether's theorem, enabling the study of signal transport along axons and synaptic connections.
• Compare phase boundaries and critical temperatures between the U(1) and classical Z(2) models to isolate quantum effects.
• Use the time evolution rule derived from the energy functional to simulate learning and recalling processes in the quantum model.

Experimental results

Research questions

• RQ1 How does replacing discrete Z(2) variables with continuous U(1) phase variables alter the phase structure of a gauged neural network model?
• RQ2 What is the role of quantum fluctuations in modifying the critical temperatures and phase boundaries compared to the classical Z(2) model?
• RQ3 Can a conserved current be defined in the quantum U(1) model, and what physical interpretation does it hold for neural signal propagation?
• RQ4 To what extent do quantum effects weaken the network's ability to learn and recall patterns compared to the classical version?
• RQ5 How does the U(1) gauge structure influence the stability and dynamics of synaptic weights during learning processes?

Key findings

• The U(1) quantum model exhibits a richer phase structure than the classical Z(2) model, including distinct Coulomb, Higgs, and confinement phases.
• The critical temperatures for phase transitions in the U(1) model are higher than those in the Z(2) model, indicating that quantum fluctuations stabilize the system against phase transitions.
• The region of the confinement phase is wider in the U(1) model compared to the Z(2) model, suggesting stronger localization of neural activity due to quantum effects.
• Quantum fluctuations reduce the network's ability to learn and recall patterns, as evidenced by the broader confinement phase and higher critical temperatures.
• A conserved gauge-invariant current $j_{x\mu}$ emerges in the U(1) model due to continuous U(1) symmetry, enabling systematic study of signal transport in neural networks.
• The model is universal in the sense of renormalization group theory, applying to a wide class of underlying microscopic quantum theories with U(1) gauge symmetry.