Bounds all around: training energy-based models with bidirectional bounds
This work addresses a key bottleneck in EBM training for researchers and practitioners in generative modeling, though it is incremental as it builds on existing minimax frameworks.
The paper tackled the difficulty of training energy-based models (EBMs) by proposing bidirectional bounds on log-likelihood to stabilize training, resulting in significantly improved stability and high-quality density estimation and sample generation.
Energy-based models (EBMs) provide an elegant framework for density estimation, but they are notoriously difficult to train. Recent work has established links to generative adversarial networks, where the EBM is trained through a minimax game with a variational value function. We propose a bidirectional bound on the EBM log-likelihood, such that we maximize a lower bound and minimize an upper bound when solving the minimax game. We link one bound to a gradient penalty that stabilizes training, thereby providing grounding for best engineering practice. To evaluate the bounds we develop a new and efficient estimator of the Jacobi-determinant of the EBM generator. We demonstrate that these developments significantly stabilize training and yield high-quality density estimation and sample generation.