RetroTune: Mitigating spurious features via retrospective fine-tuning

1. The evaluation is limited without a direct comparison with the many existing methods. A lot of related papers are missing here, see a non-comprehensive list below. Multiple lines of group inference methods are directly comparable to the proposed methods so their results should be added, and robust training methods that require group labels should be properly discussed.
2. In Section 3.4, is $a_{err}$ always zero?
3. The reported dimensions considered as extended spurious seem highly disproportionate across classes. There are in total 2048 dimensions and Table 2 shows that more than 2040 dimensions are considered as extended spurious for one class while only 4 or 20 dimensions are spurious for the other class. Can you provide insight into why this is the case and its effect on different classes?
4. Is there a specific reason for comparing only to ReBias on BiasedMNIST without other methods and not having ReBias for the other two datasets, Waterbirds and CelebA?
5. Have you considered extending your evaluations to text-based models and data?

(Sohoni et al.) No Subclass Left Behind: Fine-Grained Robustness in Coarse-Grained Classification Problems. NeurIPS, 2020.

(Creager et al.) Environment Inference for Invariant Learning. ICML, 2021.

(Zhang et al.) Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations. ICML, 2022.

(Taghanaki et al.) MaskTune: Mitigating Spurious Correlations by Forcing to Explore. NeurIPS, 2022.

(Nam et al.) Spread Spurious Attribute: Improving Worst-group Accuracy with Spurious Attribute Estimation. ICLR, 2022.

(Deng et al.) Robust Learning with Progressive Data Expansion Against Spurious Correlation. NeurIPS, 2023.

• How do you handle conflicts between different classes? For e.g., can a dimension in the latent representation be identified as a core feature for class C1 but as a spurious correlation for class C2?

• In Table 2, the number of core features is too low. Do you have any justification for why that is the case?

1. The studied setting seems to be limited. The whole paper considers OOD generalization under severe bias, without group labels. The implicit assumption makes the problem much easier to solve, while it’s unknown whether it’s the case for many practical scenarios.

2. Throughout the description of the method, there are too many heuristic designs that disable the method being applied to a general setting:

• How can the values of certain dimension reflect single invariant or spurious features? What if they are entangled together?
• Why “a spurious feature does not exist in all the training samples, and values at the corresponding dimension tend to be out-of-distribution for the inputs that do not contain the spurious feature.”? Even for a single type of background in Waterbirds, there are variations inside that is likely to change the values of the latent features. What does it mean by out-of-distribution for the inputs?
• Why “swapping core (spurious) dimensions should (should not) alter the predictions”?
• Why “masking out spurious dimensions should not alter the predictions; and masking out core dimensions should make the predictions uncertain”?
• Why “the p dimensions are very likely to be spurious” in (6)?
• How the heuristic in Sec. 3.3.2 is exactly incorporated in the whole algorithm?
• The model selection method in Sec. 3.4 is heavily dependent on the severe bias assumption.

Besides, there are no theoretical guarantees for the heuristics;

3. Experiments are limited to synthetic datasets, and only considers setting with severe bias. It’s recommended to evaluate the methods going beyond the synthetic data, such as in Wilds benchmark.

4. Results of some baselines have huge gap than those reported by the original paper. For example, in JTT, the reported worst accuracy in Waterbirds is 86.7%, significantly larger than that in Table 4.

5. Several related works are have not been discussed nor compared in the paper:

• For OOD generalization without group labels, [1] show it’s fundamentally impossible to learn the core features. How could RetroTune prevent the failure case?
• The learning of the spurious features studied in the paper also relies on tendency of optimizer to simple features[2]. There are already multiple methods to preventing the simple feature learning[3,4,]. Can these methods address the issue in the paper?
• [5,6,7] employ new methods to learn richer features, instead of spurious features, which could be severed as a more practical variant of Asgari et al., 2022 cited in the paper. Can [5,6,7] address the issue studied in the paper?
• [8,9] are also the state-of-the-art methods for the same problem studied in the paper. Can they address the issue studied in the paper?

References

[1] ZIN: When and How to Learn Invariance Without Environment Partition? NeurIPS 2021.

[2] The Pitfalls of Simplicity Bias in Neural Networks, NeurIPS 2020.

[3] Evading the Simplicity Bias: Training a Diverse Set of Models Discovers Solutions with Superior OOD Generalization, CVPR 2022.

[4] Learning an Invertible Output Mapping Can Mitigate Simplicity Bias in Neural Networks, arXiv 2022.

[5] Rich Feature Construction for the Optimization-Generalization Dilemma, ICML 2022.

[6] Learning useful representations for shifting tasks and distributions, ICML 2023.

[7] Towards Understanding Feature Learning in Out-of-Distribution Generalization, arXiv 2023.

[8] Identifying Spurious Biases Early in Training through the Lens of Simplicity Bias, arXiv 2023.

[9] Robust Learning with Progressive Data Expansion Against Spurious Correlation, arXiv 2023.

Please address the weaknesses above.