Training an Ising machine with equilibrium propagation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Physics

Background:

• Ising machines, hardware implementations of the Ising model, have historical ties to early Artificial Intelligence (AI) algorithms.
• Direct application of supervised learning methods to Ising machines has been a significant challenge, limiting their accuracy in AI tasks.
• Existing limitations hinder the full potential of Ising machines in modern AI development.

Purpose of the Study:

• To develop and demonstrate an efficient supervised training approach for Ising machines.
• To bridge the gap between Ising machine physics and essential supervised training methods for AI.
• To explore the potential of Ising machines as trainable hardware for advanced AI applications.

Main Methods:

• Utilized the Equilibrium Propagation algorithm for supervised training of Ising machines.
• Employed the D-Wave Ising machine's quantum annealing procedure for neural network training.
• Applied the approach to train a fully-connected neural network on the MNIST dataset.

Main Results:

• Achieved comparable accuracy to software-based implementations in supervised training.
• Demonstrated that Ising machine connectivity supports convolution operations.
• Successfully trained a compact convolutional network using minimal spins per neuron.

Conclusions:

• Ising machines can be efficiently trained in a supervised manner using Equilibrium Propagation.
• The D-Wave Ising machine is a viable platform for training neural networks, including convolutional networks.
• This work establishes Ising machines as a promising trainable hardware for enhancing AI and machine learning applications.