[Paper Review] Generalized Logit Adjustment: Calibrating Fine-tuned Models by Removing Label Bias in Foundation Models

Beier Zhu, Kaihua Tang+2 more|arXiv (Cornell University)|Oct 12, 2023

Domain Adaptation and Few-Shot LearningComputer Science3 citations

TL;DR

This paper proposes Generalized Logit Adjustment (GLA), a post-hoc method that removes label bias from foundation models like CLIP by estimating pre-training label distribution from downstream data alone, then ensembling the debiased zero-shot model with fine-tuned models. GLA achieves 1.5 pp accuracy gain on ImageNet and consistent improvements across few-shot and long-tailed benchmarks, outperforming standard fine-tuning and ensemble methods by being Bayes optimal under class-balanced distributions.

ABSTRACT

Foundation models like CLIP allow zero-shot transfer on various tasks without additional training data. Yet, the zero-shot performance is less competitive than a fully supervised one. Thus, to enhance the performance, fine-tuning and ensembling are also commonly adopted to better fit the downstream tasks. However, we argue that such prior work has overlooked the inherent biases in foundation models. Due to the highly imbalanced Web-scale training set, these foundation models are inevitably skewed toward frequent semantics, and thus the subsequent fine-tuning or ensembling is still biased. In this study, we systematically examine the biases in foundation models and demonstrate the efficacy of our proposed Generalized Logit Adjustment (GLA) method. Note that bias estimation in foundation models is challenging, as most pre-train data cannot be explicitly accessed like in traditional long-tailed classification tasks. To this end, GLA has an optimization-based bias estimation approach for debiasing foundation models. As our work resolves a fundamental flaw in the pre-training, the proposed GLA demonstrates significant improvements across a diverse range of tasks: it achieves 1.5 pp accuracy gains on ImageNet, an large average improvement (1.4-4.6 pp) on 11 few-shot datasets, 2.4 pp gains on long-tailed classification. Codes are in https://github.com/BeierZhu/GLA.

Motivation & Objective

• To identify and address the overlooked label bias in foundation models caused by skewed Web-scale pre-training data.
• To develop a practical method for debiasing zero-shot foundation models without access to the original pre-training dataset.
• To improve downstream performance by ensembling debiased zero-shot models with fine-tuned models in a statistically optimal way.
• To evaluate the method across diverse settings: many-shot, few-shot, and long-tailed learning, under multiple fine-tuning paradigms.

Proposed method

• GLA estimates label bias in foundation models using a constrained optimization problem that minimizes error on the downstream dataset, without requiring access to pre-training data.
• It adjusts the logit margins of the zero-shot model to correct for pre-training distribution skew, using only downstream labels and model predictions.
• The method is grounded in Bayesian decision theory, making the final ensemble a Bayes optimal classifier for class-balanced target distributions.
• GLA performs post-hoc debiasing by reweighting the zero-shot model’s logits based on estimated label priors derived from downstream data.
• The debiased zero-shot model is then ensembled with a fine-tuned model using learned fusion weights to improve robustness and accuracy.
• The approach is applied across multiple fine-tuning paradigms: end-to-end fine-tuning, linear probing, and prompt tuning.

Figure 1 : (a) Per class accuracy of CLIP-ViT/B16 on ImageNet. Class index are sorted using the estimated pre-training label prior. Curves are smoothed for better visualization. (b) Beak-down performance of different models on ImageNet. We equally divide the ImageNet classes into three subgroups, ac

Experimental results

Research questions

• RQ1 Does label bias in foundation models, induced by skewed pre-training data, degrade downstream performance, especially on rare classes?
• RQ2 Can label bias be effectively estimated and corrected without access to the original pre-training data?
• RQ3 Does ensembling a debiased zero-shot model with a fine-tuned model yield better performance than standard ensembling or fine-tuning alone?
• RQ4 How does GLA perform across diverse downstream settings, including few-shot, many-shot, and long-tailed learning?
• RQ5 Is the proposed debiasing method statistically optimal under class-balanced evaluation?

Key findings

• GLA achieves a 1.5 percentage point (pp) accuracy gain on ImageNet, significantly improving performance on tail classes.
• On 11 few-shot datasets, GLA delivers an average improvement of 1.4 to 4.6 pp, with substantial gains on rare classes.
• In long-tailed classification, GLA achieves a 2.4 pp improvement, demonstrating strong robustness to class imbalance.
• On the sketch domain of DomainNet, GLA improves zero-shot accuracy to 93.00%, outperforming both standard zero-shot (92.25%) and debiasing using pre-trained label priors (89.50%).
• The method is robust across all three fine-tuning paradigms—end-to-end, linear probing, and prompt tuning—showing consistent gains.
• GLA’s performance is superior to WiSE-FT and ProGrad, particularly on tail classes, where these methods suffer from performance degradation.

Figure 2 : Illustration of debiasing process on ImageNet validation set. (a) The original distribution of zero-shot outputs; (b) the estimated pre-train distribution $\mathbf{q}$ based on our algorithm; (c) the distribution of debiased zero-shot outputs using estimated $\mathbf{q}$ .