Learning Energy-Based Models With Adversarial Training
This work provides a competitive alternative for researchers and practitioners interested in generative modeling, particularly for image generation and translation tasks, by leveraging adversarial training for EBMs.
This paper explores using adversarial training (AT) to learn energy-based models (EBMs), demonstrating that AT learns an energy function modeling data support and is related to MCMC-based maximum likelihood. The proposed techniques achieve competitive image generation performance, stable training, suitability for image translation, and strong out-of-distribution adversarial robustness.
We study a new approach to learning energy-based models (EBMs) based on adversarial training (AT). We show that (binary) AT learns a special kind of energy function that models the support of the data distribution, and the learning process is closely related to MCMC-based maximum likelihood learning of EBMs. We further propose improved techniques for generative modeling with AT, and demonstrate that this new approach is capable of generating diverse and realistic images. Aside from having competitive image generation performance to explicit EBMs, the studied approach is stable to train, is well-suited for image translation tasks, and exhibits strong out-of-distribution adversarial robustness. Our results demonstrate the viability of the AT approach to generative modeling, suggesting that AT is a competitive alternative approach to learning EBMs.