Adaptive hyperparameter updating for training restricted Boltzmann machines on quantum annealers

Area of Science:

• Artificial Intelligence
• Quantum Computing
• Machine Learning

Background:

• Restricted Boltzmann Machines (RBMs) are used for unsupervised learning tasks like image recognition.
• Training RBMs involves optimizing likelihood over large probability spaces, often approximated during gradient-based optimization.
• Quantum annealing offers potential for more efficient probability space searching, but requires hyperparameter tuning.

Purpose of the Study:

• To propose and validate methods for estimating the inverse temperature hyperparameter for RBM training on D-Wave quantum hardware.
• To improve the efficiency and accuracy of RBM training using quantum annealing.

Main Methods:

• Estimating the inverse temperature hyperparameter by maximizing likelihood or minimizing Shannon entropy.
• Experimental validation on D-Wave hardware for an image recognition task.
• Bayesian uncertainty analysis to evaluate neural network image reconstruction errors.

Main Results:

• Both maximum likelihood and Shannon entropy minimization improved RBM training on D-Wave hardware.
• Maximum likelihood estimation resulted in over an order of magnitude lower image reconstruction error compared to manual hyperparameter optimization.
• Maximum likelihood outperformed Shannon entropy minimization in image reconstruction accuracy.

Conclusions:

• Estimating the inverse temperature hyperparameter using maximum likelihood is a superior method for training RBMs on D-Wave hardware.
• This approach significantly enhances performance in unsupervised machine learning applications like image recognition.
• The findings pave the way for more effective integration of quantum annealing in deep learning.