Improving Autoregressive Image Generation by Mitigating Gradient Bias in Softmax
Keywords: Image Generation, Autoregressive Generative Model, Softmax
Abstract: Softmax is the most commonly used probabilistic activation function in classification tasks, partly due to its tendency to over-penalize non-target classes with high prediction scores. However, this property becomes detrimental in autoregressive generation tasks, where multiple valid predictions may exist. Unlike conventional classification task, which seeks a single correct answer, autoregressive models are expected to assign high probabilities to various plausible outputs to ensure diversity in generation. However, during training, gradient bias caused by Softmax over-penalizes non-target predictions with high probabilities, limiting output diversity and hindering optimization convergence. To alleviate this, we propose Gradient Suppressed Softmax (GS-Softmax), which reduces the gradient contributions of high-probability non-target classes. Through experiments, we demonstrate that GS-Softmax improves both the diversity of generated content and optimization convergence. Code and pre-trained models will be made public.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1605
Loading