ORBIS: Open Dataset Can Rescue You From Dataset Bias

Below are several concerns related to weaknesses.

• In abstract, it is not clear how to define "open dataset", how to define "noisy information", how to define "characteristics of the given training dataset". Authors should clarify.

• The paper emphasizes the phrase "dataset bias problem" but it is not clear what it means. Is it a new problem proposed in the paper? If so, authors should clarify. Otherwise, authors should cite related papers.

• In Introduction, the paper uses "ski" as a motivational example. Concretely, the paper explains "the images of “ski” exhibit a spurious correlation with “snow,” and the snowy background is considered the bias attribute. Conversely, samples captured in a desert environment (e.g., desert skiing) would be bias conflicting samples." However, isn't it reasonable to have snow in "ski" images. In other words, "ski" is "ski", "desert ski" is a different sport. Authors should clarify.

• Figure 1 is confusing. It is not clear what pink and blue boxes mean, both are labeled as "bias-conflicting".

• In Section 2.1, the examples in the first two paragraphs are conflicting. The writing should be improved.

• Table 2 is hard to read. It has a poor caption.

• In terms of Observation 1, it is confusing to assign arbitrary labels to mined images as a compared method. Shouldn't a reasonable baseline be that of training without extra data? Moreover, what is the dataset used to report numbers for the observations?

• The paper writes "assigns labels based on cosine similarity", it is not clear why using "cosine similarity"? Why not entropy, softmax scores, etc.?

• Table 3 is hard to understand. Does it sample relevant data or just randomly sample data for pseudo labeling? It misses crucial details for understanding.

• The paper motivates the first key point, "avoid focusing solely on selecting bias conflicting samples". But it is not clear why it should "avoid solely selecting bias conflicting samples". Authors should clarify.

• The paper writes "To tackle this issue, we emply a self-supervised learning mechanism that does not rely on target labels". However, images by themselves deliver bias or spurious correlation. So why does self-supervised learning tackle the issue of data bias?

• It is not well motivated to use class centroid in self-supervised learning. Can authors explain this design choice? Why is it better than the typical contrastive loss?

• The debias loss L_{tr} is not defined. Authors should define it before using it.

• The paper writes "Furthermore, we evaluate ResNet-18 trained from scratch". Why not train ResNet50 from scratch?

• There are many typos, e.g., "bset", "for tow reasons", "pseudo-labelign", "a relevant samples", "emply". The paper needs to be polished for good readability.

W1.The writing quality is very poor.

Typos

• Page 1
  • calld -> called
  • adverasarial -> adversarial
• Page 2
  • (Figure 1) Bias-conflicting -> Bias-aligned
  • ...the number of the proportion...-> ...the number or the proportion...
  • bset -> best
• Page 3
  • calssification -> classification
  • v.s. -> vs.
• Page 4
  • consturcting -> constructing
  • implictions -> implications
  • labelign -> labeling
• Page 5
  • (Figure 2) Bias-conflicting -> Bias-aligned
  • anootations -> annotations
  • ...that do note... -> ...that do not...
• Page 6
  • $f$open -> $f_{open}$
  • ...without relying on provided or pseudo-labels. -> without relying on provided [something] or pseudo-labels.
• Page 7
  • laogirtmh -> algorithm
  • datasetinclude -> dataset include
  • Dogans and Cats -> Dogs and Cats
  • classifiaction -> classification
• Page 8
  • examels -> examples
  • plots -> plot
  • three benchmarks -> two benchmarks
• Page 9
  • philosohpy -> philosophy

Inconsistency between results and explanation.

• Page 4: 'As indicated in Table 3 semi-supervised debiasing is not a straightforward process In and can potentially degrade performance.'
• The authors described that FixMatch degrades performance, but Table 3 shows that FixMatch didn't degrade performance.

Lack of information

• On page 2, in the third contribution point, what type of labels are not required?
• What metric is reported in Table 1?
• On page 4, in Observation 1, the training instructions are missing. Did you combine the relevant samples with the original training set for training, or did you only use relevant samples?
• What metric is reported in Table 3?

Incorrect sentence

• Page 5 'As previously noted, the open dataset D_{open} does not inherently contain samples that are directly relevant to the target task.'
• I think the sentence can be revised to 'As previously noted, the open dataset D_{open} might inherently contain samples that are not directly relevant to the target task.'

W2. If the open dataset is biased, and the class-wise centroid is biased toward bias-aligned samples, could ORBIS still be helpful for debiasing if all relevant samples are bias-aligned?

W3. Is there any problem arising from the differences between the open dataset and the target dataset? What should we do if the domains of the open dataset and the target dataset are different?

W4. The applied debiasing algorithms (LfF and Disent) are limited in terms of reweighting methods and may be considered outdated. Can ORBIS effective with sampling-based debiasing method (eg. PGD), contrastive learning-based debiasing methods (ex. CNC, CDvG), mixup-based debiasing methods (ex. selecmix), and logit correction-based debiasing method (ex. LC)? PGD: Mitigating dataset bias by using per-sample gradient, ICLR 2023

• CNC: Correct-N-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlations, ICML 2022
• CDvG: Fighting Fire with Fire: Contrastive Debiasing without Bias-free Data via Generative Bias-transformation, ICML 2023
• SelecMix: Debiased Learning by Contradicting-pair Sampling, NeurIPS 2022
• LC: avoiding spurious correlations via logit correction, ICLR 2023

W5. In Table 5, I believe that the extremely high performance of ERM is due to the use of a backbone trained on the open dataset. Therefore, I speculate that the target dataset used here might be very similar to the open dataset. Consequently, it doesn't seem like the ideal target dataset for verifying the effectiveness of ORBIS.

1. contains a lot of typos and poor grammatical structure, making the paper hard to read
2. similarly, some tables lack captions and are unclear
3. experiment setup in section 3 is confusing
   • in observation 1, random samples with arbitrarily assigned labels is more like poisoning rather than taking a "natural" sample
   • in observation 2, table and description can be cleaned up to make the conclusion more clear
4. lacks ablation experiments for design design decisions made in section 4
5. observation 2 seems to be key to performance but is not explicitly used in the author's method