Joint Energy Based Model

Joint Energy-Based Models (JEMs) are hybrid discriminative-generative models aiming to simultaneously achieve high classification accuracy and generate high-quality samples. Current research focuses on improving JEM training stability, particularly addressing challenges related to sample quality in generative tasks, and enhancing their performance in various applications, including out-of-distribution detection and multimodal tasks like text-to-image generation. These advancements leverage techniques like sharpness-aware minimization, spectral normalization, and adversarial training to bridge the performance gap between JEMs and state-of-the-art methods in both discriminative and generative modeling. The resulting improvements have significant implications for robust machine learning across diverse domains.