Vision-Language Models (VLMs) learn a shared feature space for text and images, enabling the comparison of inputs of different modalities. While prior works demonstrated that VLMs organize natural language representations into regular structures encoding composite meanings, it remains unclear if compositional patterns also emerge in the visual embedding space. In this work, we investigate compositionality in the image domain, where the analysis of compositional properties is challenged by noise and sparsity of visual data. We address these problems and propose a framework, called Geodesically Decomposable Embeddings (GDE), that approximates image representations with geometry-aware compositional structures in the latent space. We demonstrate that visual embeddings of pre-trained VLMs exhibit a compositional arrangement, and evaluate the effectiveness of this property in the tasks of compositional classification and group robustness. GDE achieves stronger performance in compositional classification compared to its counterpart method that assumes linear geometry of the latent space. Notably, it is particularly effective for group robustness, where we achieve higher results than task-specific solutions. Our results indicate that VLMs can automatically develop a human-like form of compositional reasoning in the visual domain, making their underlying processes more interpretable. Code is available at https://github.com/BerasiDavide/vlm_image_compositionality.

Concept-based interpretability methods offer a lens into the internals of foundation models by decomposing their embeddings into high-level concepts. These concept representations are most useful when they are compositional, meaning that the individual concepts compose to explain the full sample. We show that existing unsupervised concept extraction methods find concepts which are not compositional. To automatically discover compositional concept representations, we identify two salient properties of such representations, and propose Compositional Concept Extraction (CCE) for finding concepts which obey these properties. We evaluate CCE on five different datasets over image and text data. Our evaluation shows that CCE finds more compositional concept representations than baselines and yields better accuracy on four downstream classification tasks. Code and data are available at https://github.com/adaminsky/compositional_concepts .

Learning composite concepts, such as \textquotedbl red car\textquotedbl , from individual examples -- like a white car representing the concept of \textquotedbl car\textquotedbl{} and a red strawberry representing the concept of \textquotedbl red\textquotedbl -- is inherently challenging. This paper introduces a novel method called Composite Concept Extractor (CoCE), which leverages techniques from traditional backdoor attacks to learn these composite concepts in a zero-shot setting, requiring only examples of individual concepts. By repurposing the trigger-based model backdooring mechanism, we create a strategic distortion in the manifold of the target object (e.g., \textquotedbl car\textquotedbl ) induced by example objects with the target property (e.g., \textquotedbl red\textquotedbl ) from objects \textquotedbl red strawberry\textquotedbl , ensuring the distortion selectively affects the target objects with the target property. Contrastive learning is then employed to further refine this distortion, and a method is formulated for detecting objects that are influenced by the distortion. Extensive experiments with in-depth analysis across different datasets demonstrate the utility and applicability of our proposed approach.

Vision-language (VL) pretrained models have achieved impressive performance on multimodal reasoning and zero-shot recognition tasks. Many of these VL models are pretrained on unlabeled image and caption pairs from the internet. In this paper, we study whether representations of primitive concepts--such as colors, shapes, or the attributes of object parts--emerge automatically within these pretrained VL models. We propose a two-step framework, Compositional Concept Mapping (CompMap), to investigate this. CompMap first asks a VL model to generate primitive concept activations with text prompts, and then learns to construct a composition model that maps the primitive concept activations (e.g. the likelihood of black tail or red wing) to composite concepts (e.g. a red-winged blackbird). We show that a composition model can be reliably learn from ground truth primitive concepts. We thus hypothesize that if primitive concepts indeed emerge in a VL pretrained model, its primitive concept activations can be used to learn a composition model similar to the one designed by experts. We propose a quantitative metric to measure the degree of similarity, and refer to the metric as the interpretability metric. We also measure the classification accuracy when using the primitive concept activations and the learned composition model to predict the composite concepts, and refer to it as the usefulness metric. Our study reveals that state-of-the-art VL pretrained models learn primitive concepts that are highly useful for fine-grained visual recognition on the CUB dataset, and compositional generalization tasks on the MIT-States dataset. However, we observe that the learned composition models have low interpretability in our qualitative analyses. Our results reveal the limitations of existing VL models, and the necessity of pretraining objectives that encourage the acquisition of primitive concepts.

Many complex problems are naturally understood in terms of symbolic concepts. For example, our concept of "cat" is related to our concepts of "ears" and "whiskers" in a non-arbitrary way. Fodor (1998) proposes one theory of concepts, which emphasizes symbolic representations related via constituency structures. Whether neural networks are consistent with such a theory is open for debate. We propose unit tests for evaluating whether a system's behavior is consistent with several key aspects of Fodor's criteria. Using a simple visual concept learning task, we evaluate several modern neural architectures against this specification. We find that models succeed on tests of groundedness, modularlity, and reusability of concepts, but that important questions about causality remain open. Resolving these will require new methods for analyzing models' internal states.