FairerCLIP: Debiasing CLIP’s Zero-Shot Predictions using Functions in RKHSs

This figure shows the label prediction step. When labels are not available for training, FairerCLIP uses cosine similarity between the text features \((X_{T})\) and image features \((X_I)\), and the text features of the sensitive group prompt \((X_{TS})\) and \(X_I\) to predict the pseudo-labels of target attributes and sensitive attributes, respectively.

This figure shows the inputs and outputs for FairerCLIP in its training stage. FairerCLIP uses representation of images and the corresponding text prompts that are constructed by target attribute \((Y)\) along with the predicted labels to find the image and text encoders, i.e., \(\mathbf{f}^*_I(.; \mathbf{\Theta_I})\) and \(\mathbf{f}^*_T(.; \mathbf{\Theta_T})\).

Shows the inference phase of FairerCLIP in which we use the trained image and text encoders to generate debiased representations from the ones generated by CLIP. The debiased representations are then used to make zero-shot predictions.

Our aim is to debias image and text features, \(X_I\) and \(X_T\), by generating representations, \(Z_I = f_I(X_I)\) and \(Z_T = f_T(X_T)\), with no or reduced statistical dependence on sensitive attribute \(S\). To measure this dependency, we need to employ a metric capable of capturing both linear and non-linear statistical dependencies.

We adopt a simplified definition of the Hilbert-Schmidt Independence Criterion (HSIC) as our dependence measure. We refer to this dependence measure between \(A\) and \(B\) as \(Dep(A, B)\). The simplification in the definition of \(Dep\) leads to some attractive properties, including:

• practical ability to capture all non-linear modes of dependencies,
• analytical tractability.

After choosing the appropriate dependence measure, we now define our objective function. Our goal is to mitigate bias in CLIP's zero-shot predictions by debiasing the underlying representations. This can be achieved by (1) reducing the information related to the sensitive attribute while (2) preserving information about the target attribute as much as possible in the pair of image-text representations and (3) keeping the image and corresponding text representations aligned with each other.

We formulate the above-mentioned learning objective through the following optimization problem.

In the above definition, the terms \(\text{Dep}(Z_I, Y)\) and \(\text{Dep}(Z_T, Y)\) contribute to maximizing the statistical dependence between the representations and the target label \(Y\), the terms \(-\tau_I\text{Dep}(Z_I, S)\) and \(-\tau_T\text{Dep}(Z_T, S)\) seek to make the representations independent of \(S\), and the term \(\tau_z\text{Dep}(Z_I, Z_T)\) ensures that the text and image features are still aligned with each other after debiasing.

In order to map the input features to the debiased representation space \((X \rightarrow Z )\), we use functions in Reproducing Kernel Hilbert Spaces (RKHSs). Our choice of RKHS is motivated by several reasons. Debiasing is inherently an optimization problem with multiple competing objectives. In such cases, optimization is the primary bottleneck rather than model expressivity. The closed-form solution afforded by our approach mitigates the optimization challenges. RKHS has nice universal approximation properties and has performance comparable to shallow MLPs while being more computationally efficient for training (Sec. 4.3 of paper) and is performant under limited data scenarios (Sec. 4.2 of paper).

A geometric illustration of the steps that FairerCLIP takes to debias the representations is shown in the above figure. In theory, the RBF kernels used in our encoder (\(\phi_I(X)\) and \(\phi_T(X)\)) map the image and text features into an infinite-dimensional space, where the samples corresponding to different target attributes are linearly separable. In the infinite-dimensional space, the encoder that optimizes the above-mentioned objective function for \(\mathbf{\Theta_I}\) and \(\mathbf{\Theta_T}\) by alternating between closed-form solvers and seeks a direction for mapping the image and text features that have low angular distance w.r.t. the direction of (1) \(Y\) labels (small \(\alpha_{IY}\) and \(\alpha_{TY}\)), (2) \(S_{\perp}\) (small \(\alpha_{IS_{\perp}}\) and \(\alpha_{TS_{\perp}}\)), and (3) the other representation (small \(\alpha_{IT}\)).

CelebA

We evaluate the performance of FairerCLIP on the CelebA dataset with high cheekbone as the target attribute and sex as the sensitive attribute.

The above table compares the performance of FairerCLIP and the baselines on the CelebA dataset with intrinsic dependency. We observe that among all baselines, Contrastive Adapter performs well and achieves appreciable EOD for the CLIP ResNet-50 model. However, most other methods seem to even amplify the bias in the original CLIP features while improving average accuracy. FairerCLIP performs the best in terms of debiasing, achieving an EOD of 0.002% and 0.005% for CLIP ResNet-50 and CLIP ViT-L/14, respectively. Overall, FairerCLIP is very effective at mitigating unfairness to a significant extent, achieving an EOD value close to zero while maintaining a high classification accuracy.

Waterbirds and CelebA

We perform numerical experiments on spurious correlation benchmarks, Waterbirds and CelebA. In Waterbirds, the target attribute is the type of bird and the sensitive attribute is the background of bird. In CelebA target attribute is blonde hair and sex is sensitive attribute. Since FairerCLIP can be learned with or without ground-truth labels, we compare it against methods from both these categories. For performance evaluation, we use three metrics: 1) Average accuracy (Avg.), 2) Worst-Group accuracy (WG), i.e., the lowest accuracy of all subgroups, and 3) Gap, which is the difference between average and worst-group accuracy.
From this table, we can make the following observations:

• On CLIP ViT-L/14, FairerCLIP has the lowest Gap and highest WG accuracy.
• For the CLIP ResNet-50, FairerCLIP outperforms the baselines in the w/o labels setting but not in the w/ label setting.

The discrepancy between the performance FairerCLIP with CLIP ViT-L/14 and CLIP ResNet-50 can be attributed to the fact that CLIP ResNet-50 features contain less information about target attributes than CLIP ViT-L/14 features, as shown in App. A.7 in the paper. Overall, the results this table indicate that FairerCLIP effectively improves the worst group’s accuracy and reduces the Gap. Notably, our approach can be applied to and is effective in both scenarios, with and without ground-truth labels. At the same time, the baselines are specialized to operate in one or the other scenario only.

FairFace

We evaluate FairerCLIP on the FairFace dataset. Here, we consider sex and race as the sensitive attributes and follow the experimental setup in Chuang et al. (2023). We use five target attributes and ten text prompts (2 prompts per attribute) unrelated to the samples' facial or sensitive attributes; we do not have access to ground-truth labels. As an example, the text prompt can be “A photo of a criminal person" or “A photo of a friendly person". All the ten specific prompts are in App. A.3. To evaluate the models, we calculate MaxSkew@1000, which assesses the maximum imbalance in certain sensitive attributes within a dataset.
As is shown in the above table, FairerCLIP outperforms the other baselines for both sensitive attributes across two different CLIP backbones.

CFD

Next, we consider a more challenging task to evaluate the data-efficiency of FairerCLIP. We use Chicago Face Database (CFD) images with attractive and sex as the target and sensitive group attributes. The former is a continuous label, which we binarize by using the mean value of all samples as a threshold. Moreover, the sex attribute is a binary label. This task presents challenges in two aspects. First, the number of samples in this dataset is very small (597 samples), which may not be sufficient for training some of the baselines. Second, the performance of the zero-shot classifier for this case shows that the features generated by the CLIP model are not well separated, rendering it difficult to correctly predict \(\hat{S}\) (see Appendix A.8).
This figure shows the results of this experiment. We first observe that all the baselines almost completely fail at mitigating the bias for the worst group. In contrast, FairerCLIP's performance is satisfyingly better, both in terms of the worst group (WG) and average (Avg) accuracy. Furthermore, the Gap is significantly lower for FairerCLIP than the other baselines.

To show the computational efficiency of FairerCLIP we report and compare the training time of FairerCLIP and other baselines in the following table.
The results show that FairerCLIP is an order of magnitude faster than most baselines and almost two orders faster than Contrastive Adapter. The underlying model for this experiment is CLIP ViT-L/14, and all the numbers are measured on the same machine.