FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in Medical Image Analysis

Additional metrics & motivation for fair learning: Thanks for the feedback. We agree that other metrics such as uniform loss or EqOpp, etc. are widely used. We emphasise that although they are common, they have also been widely criticised in the literature for being pareto inefficient and potentially violating ethical non-maleficence (Beauchamp 2003, Chen 2018, Ustun 2019, Zietlow 2022, etc). For example, it's possible to fully satisfy the EqOdds and EqOpp criteria by providing zero-accuracy for all subgroups, which would be strictly worse than the status quo. Therefore we followed the recommendation of GroupDRO (Sagawa 2019) and the recent MedFAIR (Zong 2023) and focused both our evaluation and our learning objective on the most disadvantaged subgroup metric (minAUC).

To provide more explanation, EqOpp, EqOdds and other metrics can be satisfied by both the utopia situation (100% for all subgroups) as well as pathological situations (0% for all subgroups). Meanwhile, minAUC is not satisfied by the pathological situation, and it is only fully satisfied by the utopia situation (100% for all subgroups). Thus perfect minAUC is a sufficient condition for utopia, while the others can be sufficient but are not necessary conditions.

While we prefer minAUC as a metric as explained above, for completeness we now also report other metrics - in line with requests of the other reviewers, we have decided to include Equalized Odds Difference (EOddsD) and Demographic Parity Difference (DPD). The results are in the revised paper's Table 3. We can see that FairTune does quite a good job of satisfying the EOddsD and DPD objectives, even though our algorithm optimises for minAUC. Where other methods outperform FairTune on these metrics, they are worse on both overall AUC and minAUC, thus being completely pareto dominated by FairTune.

Finally, we remark that, as discussed in Sec 4.3, etc, our overall contribution and framework is agnostic to the specific meta-objective used for FairTuning. If a user for some reason really wanted to optimize for e.g. EOddsD rather than minAUC, then this is a trivial hyper-parameter to change for FairTune, which we expect would push the results towards minimising EOddsD in favour of minAUC.

Mask differences: It is not easy to completely explain the specific masks discovered by different objectives. If there was a straightforward explanation for this, we could directly hand-design the masks rather than needing to automatically search for them. However, for example, Figure 4 shows that the minAUC objective prefers not to fine-tune the last few MLP layers, while the overall AUC objective does fine-tune them. This suggests that those later MLP layers are susceptible to overfitting to spurious correlations in the training set that can be detrimental to the disadvantaged subgroup within the testing set.

We appreciate the effort and time spent reviewing our paper, and we are grateful that you highlight that our method performs well, its design is easy to follow and is well motivated, and focuses on important topic. We respond to the weaknesses and questions next.

Binary and multi-class classification: We focus on binary classification problems as these are the ones typically considered in medical settings, and following the recent MEDFAIR benchmark (Zong 2023). However, FairTune trivially extends to multi-class classification if desired. The only required change is to replace Binary Cross Entropy loss with Multi-Class Cross Entropy in Eq (2) $L^{base}$, and adjust the AUC in $L^{fair}$ to multi-class AUC.

Challenge of PEFT: We are not sure we understand the reviewer's question. Perhaps you mean that the challenge of optimising PEFT architecture is not necessarily a fairness specific challenge? We agree that PEFT architecture optimization can be conducted with the goal of maximising conventional accuracy on a dataset, rather than maximising fairness. However, as our focus is on fairness, our contribution is to introduce a methodology that enables PEFT architecture to be optimised for fairness, and to demonstrate the empirical results that this methodology is successful. The fact that one can use related techniques to achieve other goals besides fairness does not detract from this contribution.

Details of the method: Our method tries to optimize the mask that specifies how we perform PEFT in a way to promote fairness. We optimize either 36 or 12 element masks (36 one in the main scenario - corresponding to normalization, attention and MLP layers for each of the 12 blocks of the ViT model). We use the default settings of Tree-structured Parzen Estimator (TPE) and Successive-Halving (SH) from the Optuna bilevel optimization library. TPE is a smart Bayesian Optimisation strategy for sampling architecture configurations that are the most promising to try next. While SH maintains computational efficiency by early termination of architecture configurations that are not promising. More details of their hyperparameters are now given in the Appendix, as well as the details of how we split data into train/val/test. In all these splits we have followed MEDFAIR. We have added these further details into the revised version of the paper as part of the Appendix.

Validation AUC: There seems to be a misunderstanding. Our main results in Table 1 and Table 2 all report Test AUC. We've updated the paper to make this even more unambiguous.

We appreciate the encouraging review as well as the effort and time spent reviewing our paper. We are grateful that you highlight FairTune is clearly motivated, gives strong results and aligns well with the current interest of the ML community in foundation models. We respond to the weaknesses and questions next.

Fine-tuning of self-supervised pre-trained model: Thanks for the suggestion. We have selected the masked autoencoder approach (MAE) for this evaluation and ran the experiments. The results confirm that FairTune still provides benefit when using self-supervised pre-training. Please see Table 4 in the revised paper.

Additional metrics: Thanks for the feedback. We agree that Equalized Odds Difference (EOddsD) and Demographic Parity Difference (DPD) are widely used. We emphasise that although they are common, they have also been widely criticised in the literature for being pareto inefficient and potentially violating ethical non-maleficence (Beauchamp 2003, Chen 2018, Ustun 2019, Zietlow 2022, etc). For example, it's possible to fully satisfy the EOddsD and DPD criteria by providing zero-accuracy for both subgroups, which would be strictly worse than the status quo. Therefore we followed the recommendation of GroupDRO (Sagawa 2019) and the recent MEDFAIR (Zong 2023) and focused our evaluation on the most disadvantaged subgroup metric (minAUC). This metric is not vulnerable to satisfaction by the potential pathological outcomes that can satisfy EOddsD and DPD.

While we prefer minAUC as a metric as explained above, for completeness we now also report EOddsD and DPD in the revised paper's Table 3. We can see that FairTune does quite a good job of satisfying the EOddsD and DPD objectives, even though our algorithm optimises for minAUC. Where other methods outperform fairtune on these metrics, they are worse on both overall and minAUC, thus being completely pareto dominated by FairTune.

Finally, we remark that, as discussed in Sec 4.3, etc, our overall contribution and framework is agnostic to the specific meta-objective used for FairTuning. If a user for some reason really wanted to optimize for e.g. EOddsD rather than minAUC, then this is a trivial hyper-parameter to change for FairTune, which we expect would push the results towards minimising EOddsD in favour of minAUC.

Harvard-GF3300 dataset: Thanks for the suggestion to include this dataset, we have now evaluated FairTune and all other approaches also on this dataset and added the results into our paper. The results are consistent with our results on other datasets and show the clear benefits that FairTune brings in terms of achieving superior fairness and overall performance.

We appreciate the effort and time spent reviewing our paper, and we are grateful you highlight that our work recognizes the fairness generalization gap and provides an effective workflow for adapting pre-trained models to downstream medical imaging tasks. We respond to the weaknesses and questions next.

Additional metrics: Thanks for the feedback. We agree that Equalized Odds Difference (EOddsD) and Demographic Parity Difference (DPD) are widely used. We emphasise that although they are common, they have also been widely criticised in the literature for being pareto inefficient and potentially violating ethical non-maleficence (Beauchamp 2003, Chen 2018, Ustun 2019, Zietlow 2022, etc). For example, it's possible to fully satisfy the EOddsD and DPD criteria by providing zero-accuracy for both subgroups, which would be strictly worse than the status quo. Therefore we followed the recommendation of GroupDRO (Sagawa 2019) and the recent MEDFAIR (Zong 2023), and focused our evaluation on the most disadvantaged subgroup metric (minAUC). This metric is not vulnerable to satisfaction by the potential pathological outcomes that can satisfy EOddsD and DPD.

While we prefer minAUC as a metric as explained above, for completeness we now also report EOddsD and DPD in the revised paper's Table 3. We can see that FairTune does quite a good job of satisfying the EOddsD and DPD objectives, even though our algorithm optimises for minAUC. Where other methods outperform FairTune on these metrics, they are worse on both overall and minAUC, thus being completely pareto dominated by FairTune.

Finally, we remark that, as discussed in Sec 4.3, etc., our overall contribution and framework is agnostic to the specific meta-objective used for FairTuning. If a user for some reason really wanted to optimize for e.g. EOddsD rather than minAUC, then this is a trivial hyper-parameter to change for FairTune, which we expect would push the results towards minimising EOddsD in favour of minAUC.

Code: We've now added the code as part of the supplementary material. Apologies for not including it as part of the initial submission.

Additional baselines: We now add FSCL [4] into the evaluation and our results show that while this method leads to improvements over training from scratch, it is not as good as FairTune. GroupDRO was already evaluated as part of MEDFAIR, and shown to be worse than MEDFAIR's properly tuned Full FT baseline which we also use here. Furthermore, we now also compare FairTune to another recently proposed method, FairPrune [R1], that has been shown to outperform various other methods, including AdvConf [R2], AdvRev [R3], DomainIndep [R4], and OBD [R5]. Our results clearly show that FairTune outperforms both FSCL and FairPrune.

Computational costs: FairTune is associated with larger computational costs for the fine-tuning stage as it performs optimization on how the fine-tuning is done. In practice, while the absolute cost of FairTune is larger than conventional fine-tuning, it is not prohibitively costly, eg: 14GPUh vs 48GPUh, as discussed in Sec 5. This is a one-off cost that allows us to make many fairer predictions. And as fairness is key in medical applications, the additional overhead is likely to be worth it. That said, we used a fairly simple HPO algorithm to expedite our research on fair fine-tuning, given the manageable absolute costs involved. There is scope for future extensions to reduce the computational cost, e.g., with more sophisticated gradient based search as discussed in various surveys (Liu et al, arXiv'21, Sinha et al, IEEE Trans Evo Comp'18, Hospedales et al, IEEE TPAMI'21).

In the meanwhile, to expedite large datasets, it is possible to perform FairTune search only on a subset of the dataset to keep the costs low, before actually conducting the fine-tuning on the full dataset. Subsampling for hyperparameter optimization has been used with promising results [R6, R7] and can be used to make FairTune scalable also to these settings. Random subsampling has shown promising results already [R8] and we can extend it by randomly sampling while maintaining the ratios of sensitive attributes. We have successfully used this strategy to apply FairTune to the suggested CheXpert dataset during the short rebuttal window.

CheXpert dataset: Thanks for the reminder. We have now added results on this benchmark, expediting FairTune's runtime by the simple heuristic of performing architecture search on a subset of the examples before fine-tuning the chosen architecture on the full dataset. The results are strong for FairTune and demonstrate that our approach is able to scale to large datasets by using subsampling.

Generalization: Thanks for re-iterating this point. Our motivating observation from Figure 1 aligns with the reviewer's point that there is a risk of achieving good performance and good fairness on the training data, while simultaneously achieving poor generalisation on held out data. However, while many methods optimising for fairness may lead to worse generalization on held out data, our results in Table 1 suggest that our method does not suffer from it. This is because our meta-objective for hyperparameter optimization (Eq 2) optimises for performance on held out validation data, rather than for performance on seen training data. This is an important difference between FairTune and the vast majority of existing approaches such as GroupDRO, FairPrune, FSCL and others.

References:

[R1] Yawen Wu, Dewen Zeng, Xiaowei Xu, Yiyu Shi, Jingtong Hu. Fairprune: Achieving fairness through pruning for dermatological disease diagnosis. In MICCAI, 2022.

[R2] Zhang, Brian Hu, Blake Lemoine, and Margaret Mitchell. Mitigating unwanted biases with adversarial learning. In AIES, 2018.

[R3] Eric Tzeng, Judy Hoffman, Trevor Darrell, Kate Saenko. Simultaneous deep transfer across domains and tasks. In ICCV, 2015.

[R4] Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In CVPR, 2020.

[R5] LeCun, Yann, John Denker, and Sara Solla. Optimal brain damage. In NeurIPS, 1989.

[R6] Jae-hun Shim, Kyeongbo Kong, and Suk-Ju Kang. Core-set sampling for efficient neural architecture search. In ICML Workshop on Subset Selection in ML, 2021.

[R7] Savan Visalpara, Krishnateja Killamsetty, and Rishabh Iyer. A data subset selection framework for efficient hyper-parameter tuning and automatic machine learning. In ICML Workshop on Subset Selection in ML, 2021.

[R8] Aaron Klein, Stefan Falkner, Simon Bartels, Philipp Hennig, Frank Hutter. Fast Bayesian Optimization of Machine Learning Hyperparameters on Large Datasets. In AISTATS, 2017.

We appreciate the encouraging review as well as the effort and time spent reviewing our paper. We are grateful that you highlight FairTune provides a new pathway and improvement in reducing bias in AI models and includes extensive evaluation. We respond to the weaknesses and questions next.

Computational costs: FairTune is associated with larger computational costs for the fine-tuning stage as it performs optimization on how the fine-tuning is done. In practice, while the absolute cost of FairTune is larger than conventional fine-tuning, it is not prohibitively costly, eg: 14GPUh vs 48GPUh, as discussed in Sec 5. This is a one-off cost that allows us to make many fairer predictions. And as fairness is key in medical applications, the additional overhead is likely to be worth it. That said, we used a fairly simple HPO algorithm to expedite our research on fair fine-tuning, given the manageable absolute costs involved. There is scope for future extensions to reduce the computational cost, e.g., with gradient based search as discussed in various surveys (Liu et al., arXiv'21, Sinha et al., IEEE Trans Evo Comp'18, Hospedales et al., IEEE TPAMI'21).

Scalability: For large datasets it is possible to perform FairTune architecture search without any further algorithmic improvements, simply by performing HPO on a subset of the dataset to keep the costs low, before actually conducting the fine-tuning on the full dataset. Subsampling for hyperparameter optimization has been used with promising results [R1, R2] and can be used to make FairTune scalable also to these settings. Random subsampling has shown promising results already [R3] and we can extend it by randomly sampling while maintaining the ratios of sensitive attributes. We demonstrate this is a useful strategy by using subsampling for the search on the large CheXpert dataset, where we also demonstrate excellent results.

Code: We've now added the code as part of the supplementary material. Apologies for not including it as part of the initial submission.

Evaluation on datasets with larger gaps: It is important to observe that while FairTune leads to small gaps, other baselines evaluated, for example Training from scratch and Linear readout lead to large gaps of e.g. around 10 or 20 in some setups (HAM10000 - age and Papila gender especially). As a result we already evaluate scenarios that have large gaps, and we see that FairTune is able to reduce the gaps to very small values in these cases, which is a highly desirable property to have.

Sensitivity to choice of sensitive attributes and overlapping scenarios: Our current evaluation includes 4 sets of sensitive attributes: age, gender, skin type, and race. To investigate the case of overlapping sensitive attributes, we intersected age and gender attributes using the CheXpert dataset. Our results show FairTune leads to consistent improvements even under this scenario. The details are given in the Appendix and results are shown in Table 5.

References:

[R1] Jae-hun Shim, Kyeongbo Kong, and Suk-Ju Kang. Core-set sampling for efficient neural architecture search. In ICML Workshop on Subset Selection in ML, 2021.

[R2] Savan Visalpara, Krishnateja Killamsetty, and Rishabh Iyer. A data subset selection framework for efficient hyper-parameter tuning and automatic machine learning. In ICML Workshop on Subset Selection in ML, 2021.

[R3] Aaron Klein, Stefan Falkner, Simon Bartels, Philipp Hennig, Frank Hutter. Fast Bayesian Optimization of Machine Learning Hyperparameters on Large Datasets. In AISTATS, 2017.