1. The writing is really poor. There are numerous typos, grammatical errors, and other language errors. The work requires major rewriting to be fit for a venue like ICLR. For example L98-L100, L142-L143, L145-L146, L189, L200. I suggest to the authors to carefully reread their paper and fix all of the glaring language errors, while also trying to simplify and clarify some of the more crucial sentences (for example, L178-L180 and L197-L200). The authors are almost 2 pages away from the page limit so they have a lot of space to improve the clarity of their work.
2. The description of groups in the introduction was very poor. They were, however, really neatly defined later on in Section 3.1. I expect the Introduction to be clear and accessible to people outside of the field of spurious features.
3. Figure 1 and 2 are crucial for validating the motivation behind FairDropout - poor performance on minority groups due to generalization/memorization. However, the description and analysis of the figures is lacking in clarity in my opinion. For Fig. 1, there is a mismatch between L189 and its labels, and it’s incorrectly referenced as Fig 2. In L189. For Figure 2b the description doesn’t feel clear to me. It is unclear to me how the Figures were obtained and what they represent. I expect the authors to clearly state what is the difference between minority group and worst-group (both definition-wise and practically on CellebA - which is which), and to clearly describe what they refer to as probability and how its computed.
4. The work of Maini et al. (2023) is crucial for this work, but I find the way in which it’s described (in L197-L210) very unintuitive and hard to understand. Furthermore, FairDropout is a modification of example-tied dropout from Maini et al., and yet no description nor intuition on example-tied dropout is introduced through the paper. I expect the authors to clearly state what example-tied dropout is and how it differs from their FairDropout. I also expect them to provide a clear and intuitive description of Eq. 1 so that the reader doesn’t have to read Maini et al. before reading this paper.

Summary

Based on the results and relevant literature, I believe that the idea behind the work is novel in the field of spurious features. I really like the chain of thought in this paper - starting from the problem of spurious features, to identifying overfitting on minority groups and using the recent work, that connected memorization with generalization, to propose an elegant solution. My main issue with the paper is its presentation, and I am not sure if a work that was written so poorly should get accepted to a venue like ICLR. Furthermore, due to poor comparison with the method of Maini et al., I am not sure about the novelty of the proposed methodology. Is FairDropout something new or a simple rebranding of example-tied dropout that is applied in a new field? Due to that it's hard for me to evaluate the novelty of this work - it appears novel in spurious features field, but I am unsure if it's novel in broader context.

1. It is not entirely clear to me where the novelty of the paper lies.

   • It is well-known that ERM overfits to noise in minority training examples. This is how spurious correlations are learnt and evidence for this has been previously shown already, e.g. by works like Puli et al. (2023) that the authors themselves have cited. As such, the result in Section 3.2 showing that ERM memorizes training minority samples is helpful but not novel or surprising.
   • It is also unclear to me how FairDropout is distinct from example-tied dropout as first proposed by Maini et al. (2023). I could be misunderstanding something, but besides the minor detail that dropout layers can also be applied to the projection heads, the actual implementation seems identical to example-tied dropout. As far as I can tell, it shares the same hyperparameters $p_{gen}$ and $p_{mem}$, etc.

2. The analysis in Section 3.2 is limited to CelebA. While CelebA is a benchmark for spurious correlations, it is a narrow dataset confined to faces and as such, a relatively easy dataset. I am not immediately convinced that conclusions made about memorization in CelebA can be generalized to harder datasets. At least, it would be good to show this same analysis for Waterbirds or ImageNet.

3. It is not immediately clear that deploying dropout in this manner actually fixes spurious correlations. Is there any guarantee or reasoning that generalization neurons are not also memorizing training examples, or (conversely) that memorization neurons don't learn generalizable features? The original work by Maini et al. was not designed to tackle spurious features, so I would like to some stronger evidence that this key assumption holds true. For example, at the very least, we should be able to replicate the analysis in Section 3.2 to show that zeroing out generalization/memorization neurons do/do not flip the prediction of majority/minority examples.

4. I suspect that this approach --- much like other assumption-based methods that don't require group labels (e.g. JTT, Liu et al. (2021)) --- will be very sensitive to hyperparameters. Is there: (a) any ablation for how results vary across hyperparameters, or (b) some discussion about what good hyperparameters might look like? In particular, it is a bit unclear to me how many and where the dropout layers should be placed throughout the model.

5. Finally, the actual results are a little mixed. Dropout does not seem to do well on a more challenging benchmark like Waterbirds. It has stronger results on CelebA and MultiNLI but only incremental gains on MetaShift and CXR.

[1] Puli, Aahlad Manas, et al. "Don’t blame dataset shift! shortcut learning due to gradients and cross entropy." (2023)

[2] Liu, Evan Z., et al. "Just train twice: Improving group robustness without training group information." (2021)

• The technical contribution of this paper is limited. The proposed method is a direct follow-up of the method in [1].

• The paper is poorly written. Fixing typos (e.g., Line 189: Fig. 2 -> Fig. 1, and Line 145:$p_{tr}$), grammar errors (e.g., Line 161: there no -> there is no), and format issues (e.g., different font sizes in Fig. 1) would be a good starting point. More importantly, the method is not clearly described in the paper. Specifically, how to determine what neurons represent generalizing or memorizing features is unclear.

• The proposed method needs to determine memorization neurons for each training example, which may be prohibitive when the dataset is large and there are many neurons in a model's internal layer. It would be helpful to provide the training time or an analysis on the computational complexity.

• The experiments lack a comparison with some recent methods, e.g., [2] and [3]. Moreover, the reported performance of existing methods is not standard, making it difficult to assess the effectiveness of the proposed method, in comparison with other methods such as [2] or [3]. The effectiveness of the proposed method may be strengthened if the method is evaluated in the setting of [2] or [3].

[1] Maini et al., Can neural network memorization be localized? ICML, 2023.
[2] LaBonte et al., Towards last-layer retraining for group robustness with fewer annotations, NIPS, 2023.
[3] Li et al., Bias Amplification Enhances Minority Group Performance, TMLR, 2024.

1. Missing several key baseline methods [1,2,3] in the same settings (without group labels). The reported worst-group accuracies are lower than those of these methods, which limits the contribution.
2. The observation lacks an in-depth and theoretical analysis. The observation may not always hold when it is in an under-parameterization or over-parameterization regime. See section 5.3 in [4]. To empirically validate the idea, experiments with different backbones (e.g. ResNet-101, ViT) are required or you can provide a theoretical analysis (or both).
3. Eq. 2 is hard to understand. What does the gradient subscripting $j$ mean? Section 3.2 needs to be revised for the audience to understand how certain neurons are identified.
4. There lacks an ablation study on $p_{\text{gen}}$ and $p_{\text{mem}}$.

[1] Simple and Fast Group Robustness by Automatic Feature Reweighting, ICML 2023
[2] Bias Amplification Enhances Minority Group Performance, TMLR 2024
[3] Correct-n-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlations, ICML 2022
[4] An investigation of why overparameterization exacerbates spurious correlations, ICML 2020