On the Relationship Between Double Descent of CNNs and Shape/Texture Bias Under Learning Process

2025-03-04

International Conference on Pattern Recognition

AI Summary

This work investigates the intrinsic relationship between the double-descent phenomenon and the dynamic evolution of shape/texture bias during CNN training. Using CIFAR-10/100 and Stylized-Image benchmarks, we employ gradient attribution, feature disentanglement analysis, and stage-wise model distillation to establish, for the first time, a quantitative correspondence between double-descent curves and texture-bias migration. We find that texture bias markedly intensifies at the double-descent peak, revealing a coupled mechanism among model capacity, dataset size, and inductive bias. We propose a bias-evolution analytical framework and demonstrate that early stopping effectively mitigates this bias, improving cross-domain generalization robustness by 12.3%. Our results provide novel insights into implicit regularization and generalization behavior in deep learning, linking architectural and data-driven factors to learned representational priors.

Technology Category

• Machine Learning: Deep Neural Architectures and Foundation Models
• Computer Vision: Bias, Fairness & Privacy
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Web Mining and Content Analysis: Models for Web evolution
• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Problem

• Explores relationship between CNN shape/texture bias and double descent.
• Investigates synchronized double descent/ascent of bias and test error.
• Examines double descent in CNNs without label noise conditions.

Innovation

• Explores CNN shape/texture bias relationship
• Quantifies bias during learning process
• Links bias changes to double descent