Adaptive Resolution Residual Networks

Léa Demeule, Mahtab Sandhu, Glen Berseth

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: Convolutional Neural Network, Residual Network, Adaptive Resolution, Bandwidth Robustness, Robustness, Laplacian Pyramid, Laplacian Residual, Laplacian Dropout

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

TL;DR: We develop a adaptive resolution network based on Laplacian residuals that is more robust to different image sizes.

Abstract: The majority of deep learning methods for signals assume a fixed signal resolution during training and inference, making it impractical to apply a single network at various signal resolutions. We address this shortcoming by introducing Adaptive Resolution Residual Networks (ARRNs) that implement two novel components: Laplacian residuals, which define the structure of ARRNs and allow compressing high-resolution ARRNs into low-resolution ARRNs, and Laplacian dropout, which improves the robustness of compressed ARRNs through a training augmentation. We formulate Laplacian residuals by combining the properties of standard residuals and Laplacian pyramids. Thanks to this structure, lower resolution signals require a lower number of Laplacian residuals for exact computation. This adaptation greatly reduces the computational cost of inference on lower resolution signals. This adaptation is effectively instantaneous and requires no additional training. We formulate Laplacian dropout through the converse idea that randomly lowering the number of Laplacian residuals is equivalent to randomly lowering signal resolution. We leverage this as a training augmentation that has the effect of improving the performance of the many low-resolution ARRNs that can be derived from a single high-resolution ARRN. We provide a solid theoretical grounding for the advantageous properties of ARRNs, along with a set of experiments that demonstrate these properties in practice.

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: 2234