Training Implicit Generative Models via an Invariant Statistical Loss

Authors: José Manuel de Frutos, Pablo M. Olmos, Manuel A. Vázquez, Joaquín Míguez
Venue: AISTATS 2024 · PMLR 238:2026–2034
Links: PMLR · PDF · arXiv · OpenReview · Code

Overview

We introduce the Invariant Statistical Loss (ISL), a discriminator-free objective for training implicit generative models directly from samples. Instead of learning an adversarial discriminator, ISL transforms generated samples relative to observed data and measures their discrepancy from a uniform reference distribution.

The method provides a practical sample-based alternative to adversarial training for one-dimensional generative models and extends to temporal settings, including conditional distribution learning for univariate and multivariate time series.

Relation to GANs and adversarial training

Generative adversarial networks (GANs) are a prominent class of implicit generative models. ISL addresses the same sample-based generative modeling setting while replacing the discriminator–generator min-max game with a statistical rank-based objective. This makes the method directly relevant to researchers looking for alternatives to GAN training, discriminator-free generative modeling, and more stable ways to train implicit generators without adversarial optimization.

The paper discusses the connection with adversarial generative modeling, including f-GANs, Wasserstein GANs (WGANs), and MMD-GANs. For search and comparison purposes, ISL can therefore be viewed as a non-adversarial alternative within the broader family of methods used to train GAN-like implicit generators.

Topics

invariant statistical loss · implicit generative models · generative adversarial networks (GANs) · GAN training · adversarial training · GAN alternatives · discriminator-free generative modeling · mode collapse · f-GAN · WGAN · MMD-GAN · sample-based learning · generative modeling · time series · statistical machine learning

Abstract

The canonical abstract and bibliographic record are available on the PMLR paper page.

BibTeX

@inproceedings{defrutos2024training,
  title={Training Implicit Generative Models via an Invariant Statistical Loss},
  author={de Frutos, Jos\'e Manuel and Olmos, Pablo M. and V\'azquez, Manuel A. and M\'iguez, Joaqu\'in},
  booktitle={Proceedings of the 27th International Conference on Artificial Intelligence and Statistics},
  volume={238},
  pages={2026--2034},
  year={2024},
  publisher={PMLR}
}

Recommended citation: José Manuel de Frutos, Pablo M. Olmos, Manuel A. Vázquez, and Joaquín Míguez. (2024). "Training Implicit Generative Models via an Invariant Statistical Loss." Proceedings of AISTATS, PMLR 238:2026–2034.
Download Paper