BAdd: Bias Mitigation through Bias Addition

Response to the Reviewer cPAJ

## *Responses to the questions* ### *Response to question 1* In cases where bias types or labels are unavailable, please refer to our response to reviewer comment iTN8. In cases where bias types or labels are unreliable, we present a set of experiments to illustrate the performance of BAdd under varying levels of error in the protected attribute annotations. These experiments are conducted on the UTKFace dataset, where bias is associated with the race attribute. Instead of using a bias-capturing model, we employ label projection (as in Table 15 of the manuscript), enabling the controlled injection of label errors. As shown in the table below, even with significant error levels (up to 40%—noting that 50% error corresponds to a random classifier in binary classification), BAdd achieves higher accuracy than the baseline vanilla model, which shows Unbiased and Bias-conflicting accuracy of 87.4% and 79.1%, respectively. | Bias-labels Error | Unbiased | Bias-conflicting | |-------------------|----------|------------------| | 0% | 89.68 | 88.71 | | 3% | 89.78 | 87.58 | | 5% | 89.11 | 85.21 | | 10% | 89.24 | 84.78 | | 20% | 88.58 | 81.59 | | 40% | 88.48 | 80.99 | ### *Response to question 2* In such cases, existing bias identification methods [1, 2] can be employed to infer these biases. Once identified, BAdd can be applied to mitigate them. However, it is worth mentioning that the exploration of bias identification techniques falls outside the scope of this work. For more details on this concern, please refer to our response to reviewer comment iTN8. ## *Responses to the weaknesses* For weaknesses 1 and 4, please refer to our responses to the questions. ### *Response to weaknesses 2 and 3* Some very limited added complexity occurs only in the training phase of the model. In the inference stage, the bias-capturing part of the framework is not involved at all. In addition, the fine-tuning step does not require additional resources (i.e., GPU VRAM), as it is applied after the regular training phase and involves only one trainable layer (i.e., the classification layer). For more details on the complexity of the proposed method please refer to the corresponding answer to the reviewer e7ck. [1] Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., & Shin, J. (2024). Discovering and Mitigating Visual Biases through Keyword Explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 11082-11092). [2] Zhao, Z., Kumano, S., & Yamasaki, T. (2024). Language-guided Detection and Mitigation of Unknown Dataset Bias. arXiv preprint arXiv:2406.02889.

Response to the Reviewer e7ck

1. While both BAdd and FLAC aim at learning bias-neutral representations, their approaches to bias mitigation are fundamentally different: * FLAC introduces a sampling mechanism to identify bias-aligned and bias-conflict samples. It then employs a loss function that encourages the main model to exhibit high or low similarities between samples based on the similarity levels observed in the bias-capturing representations. * In contrast, BAdd introduces a mechanism that prevents bias from being introduced to the model by injecting certain features relevant to the bias within the training phase of the model. More intuitively, FLAC (as well as most of the relevant methods) tries to penalize the bias that is already introduced in the model, while BAdd is a proactive bias mitigation method that intervenes earlier in the training process, addressing the root cause of bias propagation within the model itself. 2. Please find the additional experiments concerning 3 additional networks below. *Table: Results for UTKFace (race).* |Model|Unbiased|Bias-conflicting| |-|-|-| |ResNet18|92.24| 93.33| |EfficientNet-B0|91.89| 90.97| |Swin Transformer-Tiny|92.35| 92.12| |ViT-Base-Patch16-224|92.44|93.49| Furthermore, regarding the evaluation of the proposed method on balanced or unbiased data, we have already a relevant set of experiments presented in Table 13, where we evaluate BAdd’s performance for different bias levels, including balanced data (i.e., q=0.1). 3. To address the reviewer's comment regarding the computational complexity of the BAdd approach, we provide a detailed analysis comparing BAdd to the baseline (vanilla) model for a typical example involving a ResNet18 network, input images of size 3x224x224, and two classes for the biased attribute (similar to CelebA or UTKFace experiments). *FLOPs Analysis for Trainable Components*: - **Baseline Model**: The baseline model's computational cost is 1.818 GFLOPs. - **BAdd (Bias Feature Addition)**: The addition of bias-capturing features into the main model’s penultimate layer increases the computational complexity by 2.82e-7% (i.e., +512 FLOPs). - **BAdd Fine-Tuning Phase**: The fine-tuning process, which involves updating only the final classification layer, incurs an additional computational cost of 5.63e-7% (i.e., +1.024 FLOPs). This phase remains efficient as only a small number of parameters are updated. *FLOPs Analysis for Inference Components*: - **Projection**: If projection is employed for extracting bias features, the computational cost is only 1024 FLOPs. - **Bias-Capturing Model**: The computational cost of the bias-capturing model depends on its architecture, which in this case matches the main model. However, it is important to note that this model acts as a feature extractor, so its corresponding features need to be computed only once. *Inference Complexity*: The inference time of the BAdd approach is equivalent to the baseline, as the bias-capturing component is not involved during inference. *Number of Trainable Parameters*: Similarly, the number of trainable parameters in BAdd is the same as in the baseline model, as the additional components related to the bias features are not trainable. *Potential Overheads*: The main potential overhead arises if bias-labels are unavailable or if there is no pretrained bias-capturing model. In such cases, training a bias-capturing model from scratch would add to the overall computational effort. In summary, BAdd introduces minimal overheads during training (e.g., feature addition and fine-tuning the classification layer), while the inference complexity and number of trainable parameters remain equivalent to the baseline. 4. We appreciate the reviewer's suggestion to conduct additional ablation studies. First, we would like to highlight that Tables 11-14 involve ablation studies that sufficiently cover the main components that could potentially affect the performance of the proposed approach. Additionally, we would like to stress that in this domain, ablations involving standard hyperparameters such as batch size or optimizer types are not commonly performed due to their limited contribution to understanding the method's core behavior. Nonetheless, we have conducted a relevant experiment to assess the effect of varying the batch size and found consistent performance across different configurations (see Table below). We are open to performing further experiments if they add meaningful insights into the behavior and performance of our approach. *Table: Results for UTKFace (race).* | Batch Size | Unbiased | Bias-conflicting | |-|-|-| |32| 91.89| 93.48 | |64| 92.39| 94.01| |128| 92.24| 93.33| |256| 91.64| 93.26| |512| 90.97| 93.86|

Response to the Reviewer JPua

1. In such cases, existing bias identification methods [1, 2] can be employed to infer these biases. Once identified, BAdd can be applied to mitigate them. However, it is worth mentioning that the exploration of bias identification techniques falls outside the scope of this work. For more details on this concern, please refer to our response to reviewer iTN8. 2. BAdd is indeed tailored to visual data. However, based on the theoretical analysis provided, BAdd could potentially be applied to any context involving a feature extractor (e.g., backbone) and a classifier. While exploring its applicability to other data modalities is beyond the scope of this work, we are open to conducting further experiments on other types of data in the camera-ready version of the paper, given the limited timeline of the discussion period. 3. To clarify, the bias-capturing model is not trained concurrently with the main model; rather, it serves as a bias feature extractor. The main model is trained using standard deep learning practices, and the bias-capturing model provides features that are used to ensure that biases do not influence the final learned representations. We plan to include a relevant figure and a more detailed description in the paper to make this clearer to the reader. 4. Thank you for pointing that out. [1] Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., & Shin, J. (2024). Discovering and Mitigating Visual Biases through Keyword Explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 11082-11092). [2] Zhao, Z., Kumano, S., & Yamasaki, T. (2024). Language-guided Detection and Mitigation of Unknown Dataset Bias. arXiv preprint arXiv:2406.02889.

Response to Reviewer iTN8

1. It seems the citation [1] was referenced without the actual paper being provided. Could you please share the complete reference for this work? This will help us include it in the related works section and discuss its differences compared to our proposed method. 2. Loss spikes occur when the main model attempts to learn the correct features, as the loss temporarily increases for bias-aligned samples, which make up the majority of the training data. Injecting bias-capturing features helps prevent these loss spikes by ensuring the model continues to make correct predictions on these samples, leveraging the inherent biases in the data. This means that the bias-capturing features do not contribute to an increase in the loss; rather, they help reduce it. This allows the main model to explore other diverse features (i.e., the correct features) that further minimize the overall loss, without needing to “correct” any errors introduced by the bias-capturing module. For bias-conflicting samples, which represent only a small fraction of the training data, the bias-capturing features do contribute to the loss. Although this contribution is minimal due to the small number of such samples, it aligns with our goal of bias mitigation, as it prompts the main model to place greater emphasis on these samples, requiring additional effort to shift predictions in the correct direction despite the noise introduced by the bias features. In addition, it is important to note that “forcing zero loss” is an edge case and does not typically occur during deep learning model training. We believe that both theoretical analysis and experimental results presented in the manuscript support the described behavior but we are open to incorporate concrete suggestions if these help make our contribution clearer. 3. In a biased dataset, bias-aligned samples are by definition more (in number) than bias-conflicting samples. If $\mathcal{B}_c≫\mathcal{B}_a$, this implies that $\mathcal{B}_c$ represents the bias-aligned examples, while $\mathcal{B}_a$ represents the bias-conflicting examples. However, the reviewer may be suggesting an investigation into BAdd's performance across varying levels of dataset bias (i.e., the ratio between $\mathcal{B}_c$ and $\mathcal{B}_a$). Such an analysis is provided in the Appendix (see Table 13) and shows that BAdd demonstrates consistent performance across all the levels of data bias. 4. Indeed, BAdd requires knowledge of bias attributes, which is acknowledged as a limitation in the conclusion section of the paper. However, we would like to stress that the core contribution of the method lies in its ability to address multi-attribute biases, which remains a significant challenge in the field. To better contextualize this, we can divide the existing literature into four broad categories. First, there are bias-label aware methods for single-attribute biases, where many effective approaches exist. Second, there are bias-label unaware methods for single-attribute biases, where many approaches are available, but they tend to be less effective than those in the first category. Third, there are only a few bias-label aware methods for multi-attribute biases, which remain an open challenge as the few available bias-label aware methods have not sufficiently solved the challenge. Finally, there are bias-label unaware methods for multi-attribute biases, which currently offer no effective solutions. This categorization highlights the difficulty in handling multi-attribute biases, underscoring the importance of BAdd’s contribution. Furthermore, it should be stressed that BAdd is more flexible than typical bias-label aware methods, as it allows for the use of a bias-capturing model that can be trained on different data, thus making it more practical to apply in real-world scenarios. For instance, models processing facial images are often required not to exhibit biases with respect to certain predefined attributes, such as race, gender, or age. In such cases, it is straightforward to extract the bias features from existing pretrained models. Finally, in situations where even the bias types are unknown, existing bias identification methods can be employed to infer them [1,2]. Once the biases are identified, BAdd can be applied to mitigate them. However, the exploration of bias identification techniques is beyond the scope of this work. We hope this clarification addresses the reviewer’s concerns, and we believe that BAdd makes a meaningful contribution to the problem of multi-attribute bias mitigation in deep learning models. [1] Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., & Shin, J. (2024). Discovering and Mitigating Visual Biases through Keyword Explanation. In Proceedings of the IEEE/CVF CVPR (pp. 11082-11092). [2] Zhao, Z., Kumano, S., & Yamasaki, T. (2024). Language-guided Detection and Mitigation of Unknown Dataset Bias. arXiv preprint arXiv:2406.02889.