Published in Journal of Machine Learning Research, 27(122):1–49, 2026
A stable, discriminator-free alternative to adversarial GAN training for implicit generative models, extended to multivariate and heavy-tailed distributions with Pareto-ISL and ISL-slicing.
Paper · Project page · Code · BibTeX
Published in 29th International Conference on Artificial Intelligence and Statistics (AISTATS) — Spotlight, 2026
Dual-ISL provides a convex, likelihood-free and discriminator-free objective for neural implicit samplers, offering a stable alternative to adversarial GAN-style training.
Paper · Project page · Code · BibTeX
Published in International Conference on Machine Learning (ICML) 2026, 2026
Rank-statistic approximations of f-divergences that can be estimated directly from samples, with applications to sample-based comparison of real and generated distributions and relevance to f-GAN-style objectives.
Paper · Project page · Code · BibTeX
Published in 27th International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR 238:2026–2034, 2024
A discriminator-free, sample-based alternative to adversarial GAN training for implicit generative models, based on the Invariant Statistical Loss (ISL).
Paper · Project page · Code · BibTeX