- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper tackles long-tailed recognition by changing only the bias term of the classifier, instead of retraining the weights as decoupling methods do. During normal CE training the method periodically computes class centers. For each pair of classes it shifts the bias so that the decision boundary moves toward the head class, based on the head-class samples that sit close to the boundary. These pairwise shifts are averaged into one adjustment per class, and a regularizer keeps the bias from drifting between calibrations. Experiments on CIFAR-10-LT, CIFAR-100-LT and ImageNet-LT against eight baselines are used to support three claims, that decoupling methods overlook the bias term, that GeoBias beats single-stage methods and matches two-stage ones, and that it consistently improves tail accuracy.
Critical strengths
-
Simplicity and Clarity: The core geometric intuition is straightforward. The paper clearly and correctly illustrates how shifting the bias translates the decision boundary along the center line between two class centroids without altering the learned weight directions.
-
Ease of Integration: The method is highly plug-and-play. It integrates easily into existing training pipelines, utilizing a standard linear classifier without requiring weight normalization, data resampling, or a complex two-stage training paradigm.
Critical weaknesses
-
Flawed novelty claims regarding bias: The paper says decoupling methods overlook the bias term as a separate lever for moving the boundary, but DisAlign, which the paper cites, learns an extra bias for every class, and cRT retrains both the weights and the bias. Outside decoupling, post-hoc logit adjustment and Balanced Softmax can also be seen as per-class bias corrections, which is relevant because the conclusion extends the claim to existing methods in general. As I read it, the new part is how the bias is estimated, from the geometry of each pair of classes, and that this estimate is applied periodically during training.
-
Absence of crucial baselines: The paper does not compare with the logit adjustment of Menon et al., in its post-hoc or loss form, or with Balanced Softmax, which it does not cite. There is also no ablation that replaces the geometric target with the log ratio of the two class sizes. Published numbers suggest this comparison may not be easy, at least against Balanced Softmax. The SADE paper uses a cosine classifier for all methods. Its Softmax numbers on CIFAR-100-LT are within half a point of this paper's CE, and its Balanced Softmax numbers are higher than this paper's GeoBias numbers in all three settings it reports. The setups differ, so this is only a rough indication, but I think this is the comparison that would most strengthen the paper.
-
Substandard reproduction of baselines: Some reproduced baselines are well below their published results, which makes the comparison harder to interpret. GCL gets 39.90 on ImageNet-LT here against 54.88 in its own paper, and it is listed as single-stage although its paper trains in two stages. cRT ends up below CE on ImageNet-LT, while Kang et al. report cRT 5.7 points above joint training with the same ResNet-50 and 90-epoch setup. Separately, the claim in Table 4 that a tuned GeoBias beats the best two-stage method is based on configurations picked on the test set, and the margins over LWS are 0.23 and 0.50.
-
Severe contradictions between text and tables: A few statements in the text do not quite match the tables. On CIFAR-100-LT with ratio 200, the paper says GeoBias more than doubles GCL's tail accuracy, but 8.85 against 5.75 is about 1.5 times. On CIFAR-10-LT it is said to beat all single-stage methods consistently, while at ratio 10 it is slightly below CE, ALA and GCL. On ImageNet-LT it is said to beat GCL in overall accuracy, but GCL scores 39.90 against 39.27. The ablation reading also seems off. In Table 4, removing the anchoring regularizer improves both top-1 and tail accuracy at both ratios, whereas the text describes the regularizer as crucial for tail accuracy. I would suggest revising these sentences to match the numbers.
Questions for authors
-
What happens if the geometric target is replaced by the log ratio of the two class sizes, or if the bias is calibrated only once after training?
-
Was GCL run with only its first stage? What second-stage settings did you use for cRT and LWS, and do you have an idea why cRT is below CE on ImageNet-LT?
-
How were the default hyperparameters chosen? The paper describes no validation split apart from the test data, so I would suggest choosing both the defaults and the Table 4 settings on a validation split taken from the training set.
-
The paper mentions high training efficiency, but I could not find training times. How long does one calibration take on ImageNet-LT, where it loops over every pair of classes?