Energy-based generator matching: A neural sampler for general state space
This addresses the challenge of data-free generative modeling for researchers in machine learning, offering a modality-agnostic approach, though it appears incremental as an extension of generator matching.
The paper tackles the problem of training generative models from energy functions without data by proposing Energy-based generator matching (EGM), which enables training of arbitrary continuous-time Markov processes and generates data across modalities, achieving validation on tasks up to 100 dimensions for discrete and 20 dimensions for multimodal cases.
We propose Energy-based generator matching (EGM), a modality-agnostic approach to train generative models from energy functions in the absence of data. Extending the recently proposed generator matching, EGM enables training of arbitrary continuous-time Markov processes, e.g., diffusion, flow, and jump, and can generate data from continuous, discrete, and a mixture of two modalities. To this end, we propose estimating the generator matching loss using self-normalized importance sampling with an additional bootstrapping trick to reduce variance in the importance weight. We validate EGM on both discrete and multimodal tasks up to 100 and 20 dimensions, respectively.