Visual Analysis of the Bumpiness and Ruggedness of Residual Neural Network Landscapes

Kunyu Wu, Aichun Zhu, Guan-Nan Dong

Keywords: Fitness landscape analysis, residual neural networks, local minima

TL;DR: We propose a visualization method for visualizing neural network landscapes, which is able to reflect the features of the loss landscapes.

Abstract: Different neural network architectures result in distinct training and generalization results. For instance, deep residual neural networks are more likely to find better local minima and perform more accurate predictions than deep non-residual neural networks. However, the causes of this phenomenon still require clarification. Some works show that, for convolutional neural networks, the residual connection is beneficial to generating smooth loss landscapes. However, our visual analysis discovers opposite conclusions for MLPs. Specifically, in the XOR, Iris, and Diabetes datasets, residual MLPs tend to produce more rugged loss landscapes with stronger gradients and higher loss values than non-residual MLPs, but they can still converge to low-loss basins of attraction. In the XOR dataset, residual MLPs prefer to generate more attraction basins that are sharper and have lower loss values than non-residual MLPs. Our work advances the knowledge of residual connections.

Primary Area: visualization or interpretation of learned representations

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/2024/AuthorGuide.

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: 4360