Visual Semantic Learning via Early Stopping in Inverse Scale Space

Zhiling Zhou, Zirui Liu, Chengming Xu, Yanwei Fu, Xinwei Sun

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Keywords: Semantic Information, Inverse Scale Space, Total Varation, Sparsity, Frequency, Classification from Noisy Images

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a Total Variation regularized framework in Inverse Scale Space to learn visual semantic information, which can improve robustness, classification with noise or low resolution, etc.

Abstract:

Different levels of visual information are generally coupled in image data, thus making it hard to reverse the trend of deep learning models that learn texture bias from images. Consequently, these models are vulnerable when dealing with tasks in which semantic knowledge matters. To solve this problem, we propose an instance smoothing algorithm, in which the Total Variation (TV) regularization is enforced in a differential inclusion to generate a regularized image path from large-scale (*i.e.*, semantic information) to fine-scale (*i.e.*, detailed information). Equipped with a proper early stopping mechanism, the structural information can be disentangled from detailed ones. We then propose an efficient sparse projection method to obtain the regularized images, by exploiting the graph structure of the Total Variation matrix.
We then propose to incorporate this algorithm into neural network training, which guides the model to learn structural features in the process of training. The utility of our framework is demonstrated by improved robustness against noisy images, adversarial attacks, and low-resolution images; and better explainability via visualization and frequency analysis.

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: 4693