In this paper, we study the effect of spectral decoupling with different strengths on over-parameterized models. In the absence of weight decay, we show that the models obtained by spectral decoupling of different strengths are equivalent. When there is a small weight decay, we use the second-order Taylor expansion of the objective function to obtain an approximate solution. Analyzing the approximate solution, we find that reducing the spectral decoupling has the effect of enhancing the weight decay, which is directly equivalent in the binary classification problem. Finally, we verify our analytical conclusions through experiments.
HU Yinhan
,
GUO Tiande
,
HAN Congying
. Solutions of cross-entropy loss with spectral decoupling regularization[J]. Journal of University of Chinese Academy of Sciences, 2025
, 42(2)
: 268
-275
.
DOI: 10.7523/j.ucas.2023.071
[1] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90. DOI: 10.1145/3065386.
[2] Hendrycks D, Mu N, Cubuk E D, et al. AugMix: a simple data processing method to improve robustness and uncertainty[EB/OL]. 2019. arXiv: 1912.02781.(2020-02-17)[2023-07-15]. https://arxiv.org/abs/1912.02781.
[3] Madry A, Makelov A, Schmidt L, et al. Towards deep learning models resistant to adversarial attacks[EB/OL]. 2017. arXiv: 1706.06083.(2019-09-04)[2023-07-15]. https://arxiv.org/abs/1706.06083.
[4] Jacot A, Gabriel F, Hongler C. Neural tangent kernel: convergence and generalization in neural networks[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems. December 3-8, 2018, Montréal, Canada. New York: ACM, 2018: 8580-8589. DOI: 10.5555/3327757.3327948.
[5] Daniel S, Elad H, Shpigel N M, et al. The implicit bias of gradient descent on separable data[J]. Journal of Machine Learning Research, 2018, 19(1): 2822-2878. DOI: 10.5555/3291125.3309632.
[6] Pezeshki M, Kaba S O, Bengio Y, et al. Gradient starvation: a learning proclivity in neural networks[EB/OL]. 2020. arXiv: 2011.09468. (2021-11-24)[2023-07-15]. https://arxiv.org/abs/2011.09468.
[7] Lewkowycz A, Gur-Ari G. On the training dynamics of deep networks with L2 regularization[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems. December 6-12, 2020, Vancouver, BC, Canada. New York: ACM, 2020: 4790-4799. DOI: 10.5555/3495724.3496126.
[8] Geirhos R, Jacobsen J H, Michaelis C, et al. Shortcut learning in deep neural networks[J]. Nature Machine Intelligence, 2020, 2(11): 665-673. DOI: 10.1038/s42256-020-00257-z.
[9] Shah H, Tamuly K, Raghunathan A, et al. The pitfalls of simplicity bias in neural networks[EB/OL]. 2020. arXiv: 2006.07710. (2020-10-28)[2023-07-15]. https://arxiv.org/abs/2006.07710.
[10] Xu Z Q J, Zhang Y Y, Luo T, et al. Frequency principle: Fourier analysis sheds light on deep neural networks[EB/OL]. 2019. arXiv: 1901.06523.(2019-09-20)[2023-07-15]. https://arxiv.org/abs/1901.06523.
[11] Bartlett P L, Long P M, Lugosi G, et al. Benign overfitting in linear regression[J]. Proceedings of the National Academy of Sciences of the United States of America, 2020, 117(48): 30063-30070. DOI: 10.1073/pnas.1907378117.
[12] Hastie T, Montanari A, Rosset S, et al. Surprises in high-dimensional ridgeless least squares interpolation[J]. Annals of Statistics, 2022, 50(2): 949-986. DOI: 10.1214/21-aos2133.
[13] Shamir O. The implicit bias of benign overfitting[EB/OL]. 2022. arXiv: 2201.11489. (2022-05-29)[2023-07-15]. https://arxiv.org/abs/2201.11489.
[14] Hsu D, Muthukumar V, Xu J. On the proliferation of support vectors in high dimensions[J]. Journal of Statistical Mechanics Theory and Experiment, 2022, 2022(11): 114011. DOI: 10.1088/1742-5468/ac98a9.
[15] Vidya M, Adhyyan N, Vignesh S, et al. Classification vs regression in overparameterized regimes: does the loss function matter?[J]. Journal of Machine Learning Research, 2021, 22(1):10104-10172. DOI: 10.5555/3546258.3546480.
