Adaptive annealing schedules for BM-VAE training and generation

Develop and evaluate adaptive quantum annealing schedules for sampling from the Boltzmann-machine prior within the BM-VAE framework—beyond the fixed 5 ns diabatic quantum annealing schedule used during training and the 0.5 μs quantum annealing schedule used for unconditional and conditional generation—to determine whether such adaptation improves both training convergence and generation quality.

Background

The paper introduces a variational autoencoder with a general Boltzmann-machine prior and operates a D-Wave Advantage2 annealer in three modes: DQA for training (fast schedule, aiming at β≈1), QA for unconditional generation (slower schedule), and c-QA for conditional generation (slower schedule with external bias fields). These modes in the study use fixed default annealing times (5 ns for training; 0.5 μs for generation).

In the Discussion, the authors state that several directions remain open, explicitly identifying the design of adaptive annealing schedules as a potential way to further improve both training and generation quality. This raises the concrete problem of creating and assessing adaptive (e.g., data- or epoch-dependent) schedules that could outperform the fixed defaults within the same BM-VAE setup.