Vision language models gain clearer concept level understanding
A General Concept-based Decomposition for Vision-Language Embeddings
Computer Vision and Pattern Recognition
Summary
Vision-language models try to understand images and describe them like people would. But it's hard to know if these models truly think about the parts of an image the way humans do. The authors created a method called GRACE that breaks down how these models understand images into meaningful concepts that make sense globally, not just for one example. This helps explain what the model sees in a way closer to human perception, without hurting how well the model works.
What this means in practice
- •For machine learning engineers: Enhance the interpretability of vision-language models by revealing concept-level contributions within multi-modal embeddings without performance loss.
- •For video content analysts: Improve semantic understanding and description of video data by decomposing visual embeddings into human-aligned conceptual parts across different tasks and domains.
Authors
Simone Alberto Peirone, Ortal Senouf, Francesca Pistilli, Giuseppe Averta, Pascal Frossard
Abstract
Vision-Language Models are designed to capture the compositional structure of the world through multi-modal alignment; however, it remains unclear whether their representations truly align with the way a human would naturally describe the same visual input. Prior works attempt to decompose visual embeddings into their concept-level contributions, but their instance-level optimization is inherently local and does not capture the model's global semantic structure, which is essential for compositional generalization across different inputs, tasks, and domains. We introduce GRACE, a method that learns this concept-based structure at the feature space level, producing semantic and general concept decompositions. To achieve this, we propose a novel training objective that preserves the geometric relations of different samples in the concept space, enabling a better understanding of how concepts compose and relate to each other. We validate GRACE on several image and video benchmarks, showing that its concept decomposition is aligned with human perception and achieves faithful decompositions across different tasks and domains. With minimal overhead, GRACE makes existing Vision-Language Models more interpretable, while preserving most of their original performance.