On the Unreasonable Effectiveness of Last-layer Retraining

2025/12/01 by John Hill, Hill, John C., Tyler LaBonte +5
Computer Science · #Advanced Neural Network Applications #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Stochastic Gradient Optimization Techniques

paper · pdf · doi:10.48550/arxiv.2512.01766

openalex publication_date 2025/12/01 · openalex created_date 2025/12/03 · openalex updated_date 2026/07/28

Abstract

Last-layer retraining (LLR) methods -- wherein the last layer of a neural network is reinitialized and retrained on a held-out set following ERM training -- have garnered interest as an efficient approach to rectify dependence on spurious correlations and improve performance on minority groups. Surprisingly, LLR has been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set. We initially hypothesize that this ``unreasonable effectiveness'' of LLR is explained by its ability to mitigate neural collapse through the held-out set, resulting in the implicit bias of gradient descent benefiting robustness. Our empirical investigation does not support this hypothesis. Instead, we present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set. We conclude by showing how the recent algorithms CB-LLR and AFR perform implicit group-balancing to elicit a robustness improvement.

Citations

• Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions
• Collapsed Language Models Promote Fairness
• The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations
• Navigate Beyond Shortcuts: Debiased Learning through the Lens of Neural Collapse
• Robustness to Subpopulation Shift with Domain Label Noise via Regularized Annotation of Domains
• Simplifying Neural Network Training Under Class Imbalance
• TLDR: Text Based Last-layer Retraining for Debiasing Image Classifiers
• Neural Collapse for Unconstrained Feature Model under Cross-entropy Loss with Imbalanced Data
• Towards Last-layer Retraining for Group Robustness with Fewer Annotations
• Neural Collapse in Deep Linear Networks: From Balanced to Imbalanced Data
• Freeze then Train: Towards Provable Representation Learning under Spurious Correlations and Feature Noise
• On Feature Learning in the Presence of Spurious Correlations
• MaskTune: Mitigating Spurious Correlations by Forcing to Explore
• Imbalance Trouble: Revisiting Neural-Collapse Geometry
• Neural Collapse: A Review on Modelling Principles and Generalization
• Why does Throwing Away Data Improve Worst-Group Error?
• Improved Group Robustness via Classifier Retraining on Independent Splits
• Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
• Correct-N-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlations
• Limitations of Neural Collapse for Understanding Generalization in Deep Learning
• Simple data balancing achieves competitive worst-group-accuracy
• Just Train Twice: Improving Group Robustness without Training Group\n Information
• Neural Collapse Under MSE Loss: Proximity to and Dynamics on the Central Path
• WILDS: A Benchmark of in-the-Wild Distribution Shifts
• Noise or Signal: The Role of Image Backgrounds in Object Recognition
• The Pitfalls of Simplicity Bias in Neural Networks
• An Investigation of Why Overparameterization Exacerbates Spurious Correlations
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization
• Distributionally Robust Language Modeling
• Nuanced Metrics for Measuring Unintended Bias with Real Data for Text\n Classification
• Recognition in Terra Incognita
• The Implicit Bias of Gradient Descent on Separable Data
• A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference
• Fair prediction with disparate impact: A study of bias in recidivism\n prediction instruments
• Deep Residual Learning for Image Recognition
• Aligning Books and Movies: Towards Story-like Visual Explanations by Watching Movies and Reading Books
• Deep Learning Face Attributes in the Wild
• ImageNet Large Scale Visual Recognition Challenge
• Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

Related

• Towards Last-layer Retraining for Group Robustness with Fewer Annotations
• Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining
• Theoretical Guarantees of Data Augmented Last Layer Retraining Methods
• Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
• Is Last Layer Re-Training Truly Sufficient for Robustness to Spurious Correlations?