1: \begin{abstract}

2: Maximum likelihood (ML) learning for energy-based models (EBMs)

3: is challenging, partly due to non-convergence of Markov chain Monte Carlo.

4: % because Markov chain Monte Carlo required may suffer non-convergence for complex, multimodal distributions.

5: Several variations of ML learning have been proposed, but existing methods all fail to achieve

6: both post-training image generation and proper density estimation.

7: We propose to introduce diffusion data and learn a joint EBM, called diffusion assisted-EBMs, through persistent training

8: (i.e., using persistent contrastive divergence) with an enhanced sampling algorithm to properly sample from complex, multimodal distributions.

9: We present results from a 2D illustrative experiment and image experiments and demonstrate that,

10: for the first time for image data, persistently trained EBMs can {\it simultaneously} achieve

11: long-run stability, post-training image generation, and superior out-of-distribution detection.

12: \end{abstract}

13: