3VL: using Trees to teach Vision & Language models compositional concepts

Vision-Language models (VLMs) have proved effective at aligning image and text representations, producing superior zero-shot results when transferred to many downstream tasks. However, these representations suffer some key shortcomings in Compositional Language Concepts (CLC) understanding such as recognizing objects' attributes, states, and relations between different objects. Moreover, VLMs typically have poor interpretability, making it challenging to debug and mitigate compositional-understanding failures. In this work, we introduce the Tree-augmented Vision-Language (3VL) model architecture and training technique accompanied by our proposed Anchor inference method and Differential Relevance (DiRe) interpretability tool. By expanding the text of an arbitrary image-text pair into a hierarchical tree structure using language analysis tools, 3VL allows inducing this structure into the visual representation learned by the model, enhancing its interpretability and compositional reasoning. Additionally, we show how Anchor, a simple technique for text unification, can be employed to filter nuisance factors while increasing CLC understanding performance, e.g., on the fundamental VL-Checklist benchmark. We also exhibit how DiRe, which performs a differential comparison between VLM relevancy maps, enables us to generate compelling visualizations of the reasons for a model's success or failure.