CLIP Exhibits Improved Compositional Generalization Through Representation Disentanglement

Point 5: While prior work has explored the relationship between training data diversity and compositional generalization, our study provides new analysis quantifying how diversity in the training data leads to improved disentanglement in model representations, thereby directly enabling better out-of-distribution (OoD) generalization.

Specifically, we conducted the first analysis (Section 3.3.1) that examines the impact of caption diversity and quality in training datasets on the mutual information (MI) between object and attribute tokens. We found that lower NMI indicates higher disentanglement between the two, which can improve the compositional generalization.

Building on this, we propose a way to measure the disentanglement between objects and attributes in the representation spaces of CLIP text and image encoders (Section 3.3.2). To our knowledge, this disentanglement has not been explored or validated before in CLIP models. We demonstrate models with more disentangled representations achieve higher performance on our new compositional OoD benchmark. We also provided evidence that the disentanglement is much stronger on the text representation, and provided some theoretical insights on how the disentanglement could be induced on the image representation from the text.

Furthermore, we provide additional insights into CLIP disentanglement beyond our benchmark by evaluating on 3D Shapes datasets. Our work is the first to extensively examine and validate the connection between disentangled representations and compositional generalization performance of CLIP models.

Overall, our analysis and experiments on quantifying disentanglement and its impact on compositional generalization enhance the understanding of why and how CLIP models are able to generalize compositionally. We extend current knowledge by empirically demonstrating that diversity promotes compositionality through inducing disentangled representations.

Question: Since DINO-v2 and BEiT-v2 are primarily image encoders without an accompanying text encoder, they fall outside the scope of our evaluation criteria. Most of our experimental design required models with text encoding capabilities to comprehensively assess and compare their performance.

Dear reviewer,

While we admit some shared concerns between the reviewers at the beginning, we provided several clarifications in our response that might have been overlooked. For instance, we provided the detailed information on how the attributes are selected in the rebuttal, which is NOT further criticized by any of the reviewers. Also, could you please mention which compositionality claims in the paper is unsupported or has issues? We provided several further evidences along the disengagement of representations results in our response. Human evaluation has also been addressed in our response and we believe that we clearly mentioned the criteria that are used to make the assessment (e.g. object recognition and attribute visibility). We kindly ask the reviewer to guide us on which aspect of this assessment is still unclear to him/her.

Dear Reviewer PAYB:

Thanks for the comments on the authors' rebuttal. The authors have posted additional clarification on the focus of their paper and the dataset design details. Please check if you have overlooked any details here:

1. Addressing Shared Concerns and Providing Clarifications: While we acknowledge the shared concerns initially raised by the reviewers, we believe that our comprehensive clarifications provided in the response might have been overlooked. Specifically, we detailed how the attributes for our study were selected, which was a key point of concern. Notably, this aspect of our response did not receive further criticism from any of the reviewers, suggesting that our clarification was satisfactory.

2. Emphasizing the Primary Contribution of Our Paper: We would like to reemphasize that the primary focus of our paper is on how increased disentanglement in CLIP representations aids in generalizing compositional Out-of-Distribution scenarios. This is substantiated by experiments on two datasets, including our custom-generated dataset and the 3D shape dataset. The consistent results strongly support our hypothesis, and we believe this critical aspect may have been somewhat overshadowed in the discussions about dataset design.

4.The Authenticity of Our Dataset Design: In order to ensure the validity of our dataset design, we adhered to the following key principles:

Comprehensive and Systematic Attribute and Object Selection: We meticulously selected a diverse range of objects (representing entire ImageNet classes) and comprehensive attributes. This approach aimed to ensure inclusivity., and systematic representation.

Ensuring Originality: We took care to thoroughly validate the uniqueness of the captions in our dataset, guaranteeing no overlap with any of the captions from current open-source training datasets utilized for CLIP models. Any captions discovered to match those in the training sets were eliminated from our dataset.

Comprehensive Human Evaluation: To guarantee the reliability of our dataset, we performed exhaustive human evaluations. Two separate reviewers stringently examined the captions to validate they accurately portrayed the corresponding images. Any instances of disagreement were meticulously re-evaluated to ensure consistency and accuracy across the dataset. Through this comprehensive review process, we upheld precision and fidelity in our data.

---

Please review their response and adjust you scores if the response address your concerns.

AC

Dear Reviewer PAYB:

The authors have the following comments to address your concerns. Please let me know if they addressed your concerns. Thanks,

AC

---------------Comments from the authors:

#Clarification on CLIP Models and Datasets:

You mentioned that the claim regarding the elimination of captions from training sets applies only to OpenCLIP and not all CLIP models. It's important to clarify that this approach was indeed applied to all models evaluated in our study. We thoroughly investigated and discovered datasets for each model, including CC12m, YFCC15m, LAION, and Datacomp. The only exception was the OpenAI model, which uses a private dataset. It's essential to note that OpenCLIP is not a singular model but a comprehensive project and repository that amalgamates various available CLIP models. OpenCLIP stands out as one of the most popular repositories on GitHub within the field of CLIP and VLM models. Many research papers and teams dedicated to CLIP utilize this repository as a central hub, contributing various CLIP models, and the latest advancements are consistently added to enrich this collaborative resource. Repository link : https://github.com/mlfoundations/open_clip

#Revisions and Supplementary Material:

Regarding the lack of revisions and updates to supplementary materials, I assure you that we have made substantive revisions based on the discussions between reviewers and authors. These revisions will be more apparent in the final version of the paper. Please see the next comment, where we outlined the items to be added to the Appendix. We understand the importance of detailing our methodology for selecting attributes and ensuring transparency. Therefore, we have included a comprehensive explanation in the Appendix of the final paper. This addition aims to address the major points raised during the discussion period and provide clarity on our methods. Please see the next comment for further details.

Point 4: In the context of Normalized Mutual Information (NMI) in our experiments, 'X' represents the entire set of attributes that we utilized in generating our dataset, encompassing various descriptive elements such as texture. 'Y', on the other hand, corresponds to the set of objects used in generating our dataset, including a range of specific items or entities. These definitions of 'X' and 'Y' enable us to analyze the relationship between attributes and objects in our dataset as learned by the CLIP model.

\\begin{aligned} & \\mathrm{MI}(\\mathrm{X} ; \\mathrm{Y})=E_y[\\mathrm{H}(\\mathrm{X})-\\mathrm{H}(\\mathrm{X} \\mid \\mathrm{Y}=\\mathrm{y})] \\\\ & \\mathrm{H}(\\mathrm{X})=-\\sum_{i=1}^n p_i^* \\log \\left(p_i\\right) \\\\ & p\\left(x_i\\right)=\\frac{\\# \\text { captions that have } x_i}{\\sum_{j=1}^n \\text { \\#captions that have } x_j} \\\\ & p(x, y)=\\frac{\\# \\text { captions that have } x \\text { and } y}{\\sum_{x y} \\sum_y \\text { captions that have } x_i \\text { and } y_j} \\\\ & \\mathrm{H}(\\mathrm{X} \\mid \\mathrm{Y}=\\mathrm{y})=-\\sum_x p(x, y) * \\log \\left(\\frac{p(x, y)}{p(y)}\\right) \\\\ & \\mathrm{NMI}(\\mathrm{X} ; \\mathrm{Y})=\\frac{2 * M I(X ; Y)}{[H(X)+H(Y)]} \\end{aligned}

Point 5: As mentioned in Section 3.3.2, we can consider two super latent factors corresponding to attributes and objects. For each vector of factors, we choose a image associated with those factors. We then calculate the embedding for that image. This gives us a set of representations and corresponding latent factors. Using these embeddings and factors, we can calculate different disentanglement metrics. To assess the randomness effects, we repeated this experiment multiple times, and observed that the results show negligible variance. We also observed that if instead of random selection, we calculate embeddings of all images per relation and take the average, the results hold. We note that this is because the images associated with a certain relation have similar embeddings. For more details on how each metric is calculated, please refer to [2].

Point 6: As detailed in Section 6.2 of the Appendix, we utilized the entire spectrum of 1000 classes from the ImageNet dataset.

Point 7: We appreciate the opportunity to provide additional details on the methodology used to generate our dataset. As mentioned in Section 6.2 and illustrated in Figure 3, we employed a simple template structure, [ATTRIBUTE] [OBJECT], to create compositional images via a text-to-image model. This approach allowed us to systematically combine various attributes with objects to enrich the diversity and complexity of the images. Also this approach was used in previous works [3,4]. For concrete examples of these prompts, one may refer to Figure 2 in the main text, which showcases sample image captions generated using this template. Furthermore, a more extensive collection of prompts can be found in Figure 8 of the Appendix.

Point 8: Initially, we compiled an extensive list of adjectives using various sources, including linguistic websites and large language models like ChatGPT, to ensure a broad and inclusive initial selection. The primary list contained 50 attributes. Through iterative experimentation, we generated preliminary images for a range of objects paired with these adjectives. This practical evaluation was crucial as it allowed us to observe the representation of certain attributes within the images. We found that some adjectives, such as ‘fast’, did not visually translate well into the generated images. Consequently, these were omitted from the final list. The refinement process led us to a curated set of 30 adjectives that were visually discernible and thus suitable for our image generation objective. The full methodology for selecting attributes will be detailed in the Appendix.

Point 9: We have to clarify that by 'after this process ...,' we meant to report the number of generated images after the final step, which is the human inspection. Here we got total of 50k images from 12k combinations that passed the human evaluation criteria. The human inspection was conducted in two stages: First, before generating any images, we filtered the initial list of 30,000 combinations by searching for these combinations in the captions of training datasets and removing any matches, to ensure the remaining combinations were novel and feasible to depict visually. Second, after generating images for the remaining novel combinations using the text-to-image model, further human evaluation was necessary to eliminate unsatisfactory images where either the object was not clearly rendered or the intended attribute was not distinctly perceptible. Through this two-step refinement process, we narrowed down the dataset from 30,000 initial combinations to a final curated collection of around 12,000 novel (attribute,object) combinations comprising 50,000 images that passed the human evaluation criteria.

I thank the authors for the detailed and thorough response, which addresses several of my concerns. Please find my comments below-

• Point 1: In this case, the following statement needs to be removed from the Introduction - "For instance, the training dataset of the OpenAI CLIP has not been released, which makes designing a test set that has a truly different distribution from the training one challenging. We aim to address these issues in this paper, by ... " This is because, the construction of ImageNet-AO assumes access to the training dataset of CLIP, while this statement makes it seem like access is not needed.
• Point 2: It is possible that synonyms of object-attribute pairs exist as captions to similar images. Although the exact pair may not exist in the training dataset, a more robust means of ensuring that a similar image-caption combination does not exist in the training dataset is required. For example, finding the k nearest neighbors of the caption and image individually, and using human inspection to ensure the k nearest captions and k nearest images are not similar to them. Since the training dataset is very large scale, we cannot rely on the combination being unlikely to exist in the dataset, although it seems to be the case.
• Point 3: Could the authors clarify whether the first row in the table presented in the rebuttal uses text encoder or image encoder? Also, could the authors share the conclusions from the table?
• Point 9: For human evaluation, it is important to also include details on how this was conducted, and how many people participated in the study.

I update my score to 5 based on the rebuttal.

Point2 : Table 1 in Section 3.3.1 now includes NMI values for various subsets of the LAION dataset. It is worth noting that these different LAION subsets originate from the same source but undergo different filtering processes. The filtering appears to alter the token distribution, which in turn affects the NMI values. However, NMI has no direct relationship to the dataset size. For instance, as the results indicate, despite its smaller size, LAION 400m shows a bit higher composition diversity compared to that the LAION 2B.

Table1: Normalized Mutual Information between the attributes and objects calculated for captions of some CLIP training sets. The domain of these random variables are defined based on the compositions present in our generated dataset.

Additionally, Section 4.1 now contains expanded results, including additional models in the updated version of Table 2.

Table2: Number of common dimensions across factors and switching dimensions for color manipulation in 3D Shapes dataset

Point 3: In Section 4.1, when we refer to 'dimensions,' we are speaking of the individual elements within the embedding vector generated by the CLIP model. To clarify the concept of 'switching dimensions,' we conducted experiments where these individual elements are systematically altered to assess their impact on the performance of the model. By switching, we mean replacing the values in these dimensions by those of the samples that have a different level of attributes. This process helps us understand which dimensions of the embedding vector correspond to specific attributes in the data.

Regarding the image retrieval task discussed in Section 4.2, we propose that a more decomposable and disentangled representation leads to better retrieval performance. This is because when the embedding of an image is decomposable, adding the embedding of an attribute (like a color or shape) to the image’s embedding results in an alteration of only the intended attribute in the resultant image, without affecting other characteristics. Conversely, if the embedding is entangled, the same operation would inadvertently alter the entire image, not just the intended attribute. Therefore, the ability to perform precise modifications through embedding manipulation is indicative of a disentangled representation, which, in turn, suggests higher accuracy in retrieval tasks.

In our additional experiment using the 3D Shapes dataset, we aim to correlate the model's disentanglement metrics with its Out-of-Distribution (OoD) generalization performance. We employ a classifier trained to recognize common features across different images, such as color. By identifying which dimensions of the embedding vector are significant for such features, we gain insights into the disentanglement of the embeddings. Our findings suggest that in higher-performing CLIP models on OoD benchmarks, fewer dimensions need to be switched to achieve an embedding that aligns with a modified attribute, thus reinforcing the connection between disentanglement and OoD performance.

We have included additional diagrams in the revised manuscript to visually demonstrate the processes and outcomes of these experiments, specifically experiment 4.1, for greater clarity.

We hope that this explanation resolves the ambiguities and look forward to further comments and suggestions.