Diabatic quantum annealing for training energy-based generative models.

Area of Science:

• Quantum Computing
• Machine Learning
• Statistical Physics

Background:

• Energy-based models require unbiased Boltzmann samples for training.
• Classical sampling methods (e.g., Markov chain Monte Carlo) are slow and produce correlated samples, hindering large-scale training.

Purpose of the Study:

• To address the bottleneck in training energy-based models by utilizing quantum annealing for Boltzmann sampling.
• To enable faster convergence and reduced validation error in restricted Boltzmann machines (RBMs).

Main Methods:

• Applied the analytic relation between annealing schedules and effective inverse temperature in diabatic quantum annealing.
• Implemented sampling on a quantum annealer to obtain temperature-controlled Boltzmann samples.
• Developed an analytical rescaling method to mitigate temperature misalignment noise in analog quantum computers.

Main Results:

• Achieved faster convergence and lower validation error in RBM training compared to classical sampling.
• Demonstrated that quantum annealers can serve as practical Boltzmann samplers.
• Showcased a method where model connectivity is determined by qubit connectivity, shifting computational complexity to hardware.

Conclusions:

• Quantum annealing offers a viable alternative to classical sampling for training energy-based models.
• The proposed rescaling method enhances the practicality of quantum annealers for Boltzmann sampling.
• This approach extends to fully connected Boltzmann machines, opening new possibilities beyond classical training.