Embedding Principle in Depth for the Loss Landscape Analysis of Deep Neural Networks

Embedding Principle in Depth for the Loss Landscape Analysis of Deep Neural Networks

Year:    2024

Author:    Zhiwei Bai, Tao Luo, Zhi-Qin John Xu, Yaoyu Zhang

CSIAM Transactions on Applied Mathematics, Vol. 5 (2024), Iss. 2 : pp. 350–389

Abstract

In this work, we delve into the relationship between deep and shallow neural networks (NNs), focusing on the critical points of their loss landscapes. We discover an embedding principle in depth that loss landscape of an NN “contains” all critical points of the loss landscapes for shallower NNs. The key tool for our discovery is the critical lifting that maps any critical point of a network to critical manifolds of any deeper network while preserving the outputs. To investigate the practical implications of this principle, we conduct a series of numerical experiments. The results confirm that deep networks do encounter these lifted critical points during training, leading to similar training dynamics across varying network depths. We provide theoretical and empirical evidence that through the lifting operation, the lifted critical points exhibit increased degeneracy. This principle also provides insights into the optimization benefits of batch normalization and larger datasets, and enables practical applications like network layer pruning. Overall, our discovery of the embedding principle in depth uncovers the depth-wise hierarchical structure of deep learning loss landscape, which serves as a solid foundation for the further study about the role of depth for DNNs.

You do not have full access to this article.

Already a Subscriber? Sign in as an individual or via your institution

Journal Article Details

Publisher Name:    Global Science Press

Language:    English

DOI:    https://doi.org/10.4208/csiam-am.SO-2023-0020

CSIAM Transactions on Applied Mathematics, Vol. 5 (2024), Iss. 2 : pp. 350–389

Published online:    2024-01

AMS Subject Headings:    Global Science Press

Copyright:    COPYRIGHT: © Global Science Press

Pages:    40

Keywords:    Deep learning loss landscape embedding principle.

Author Details

Zhiwei Bai

Tao Luo

Zhi-Qin John Xu

Yaoyu Zhang

  1. Implicit Regularization of Dropout

    Zhang, Zhongwang

    Xu, Zhi-Qin John

    IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 46 (2024), Iss. 6 P.4206

    https://doi.org/10.1109/TPAMI.2024.3357172 [Citations: 2]