TL;DR
This paper introduces VISaGE, a new dataset designed to evaluate how well Vision-Language Models (VLMs) handle conceptual knowledge, particularly when faced with exceptions to general rules. They found that VLMs struggle to reconcile their learned conceptual priors with atypical visual instances, often prioritizing visual-textual congruence over semantic accuracy.
What Problem Does This Paper Address?
Vision-Language Models (VLMs) are trained on vast amounts of data, learning to associate visual information with textual descriptions. This allows them to build generalized conceptual knowledge. However, when these models are presented with an image that is an exception to a general rule (e.g., a cat with only three legs when the general concept is "cats have four legs"), their performance can degrade. This happens because VLMs are influenced by two competing "priors":
- Pragmatic Prior: The model assumes that the text and image provided are congruent and relevant to each other, a bias learned from typical training data.
- Semantic Prior: The model relies on its learned generalized knowledge about categories (e.g., "cats have four legs").
When an image is an exception, these two priors can conflict. The paper aims to understand how VLMs navigate this tension and how robust their conceptual understanding is when faced with visual variations within a category. Existing evaluations often focus on typical instances or simple attribute recognition, leaving the understanding of exceptions and their impact on conceptual reasoning underexplored.
Methods
The researchers designed two main experimental setups to probe VLMs:
- Conceptual Attribute Prediction (Exp. ①): This experiment tests whether VLMs can access general conceptual knowledge even when presented with atypical images. They query models about generic statements (e.g., "Do lions have manes?") using both typical and exceptional images, and also use exception-specific terms (e.g., "Do lionesses have manes?").
- Instance Attribute Recognition (Exp. ②): This experiment assesses VLMs' ability to recognize specific attributes of individual instances, even when those instances are exceptions. They use queries like "Does this lion have a mane?" with both typical and exceptional images, and also provide explicit cues using exception names.
To understand why models behave as they do, they also employed Shapley Value analysis (MM-SHAP). This technique helps quantify the contribution of image versus text inputs to the model's predictions, revealing which modality the model relies on more heavily under different conditions.
Datasets
The core contribution of this paper is the introduction of the VISaGE (Visual Generics and Exceptions) dataset.
- VISaGE Dataset: This dataset consists of conceptual categories paired with images of both typical and exceptional instances. It includes:
- Category-attribute pairs (e.g., "cats have four legs").
- Exceptions to these norms (e.g., "tripod cats" as an exception to "cats have four legs").
- Images representing both the typical instances and the exceptional instances.
- The dataset was constructed by extracting norms from existing resources (XCSLB, McRae norms) and generating exceptions using LLM prompting, followed by extensive manual validation. It contains 1601 exceptional image examples across 437 subcategories, derived from 296 category-attribute relations.
The paper also mentions using the THINGS dataset for typical object instances.
Key Insights and Ideas
- Pragmatic Prior Dominance: VLMs tend to prioritize the visual-textual congruence learned during training (pragmatic prior) over their semantic knowledge when faced with incongruent inputs. This means an atypical image can mislead the model even when it "knows" the general rule.
- Conceptual Understanding is Fragile: When asked about general concepts (Exp. ①), VLMs' accuracy drops significantly when presented with incongruent (exceptional) images, highlighting that their conceptual representations are not robust to within-category variation.
- Instance Recognition Also Affected: Even when asked about specific instances (Exp. ②), models often rely on semantic priors (general knowledge) rather than visual features, leading to errors when the instance is exceptional.
- Explicit Cues Help, But Not Always: Using exception-specific terms (e.g., "lioness" instead of "lion") can improve performance, but the benefit is not always substantial, suggesting that simply naming an exception doesn't fully overcome the model's biases.
- Image Contribution Varies: Shapley value analysis shows that while models generally attend to images, they recruit visual information more when text and image are incongruent or atypical, suggesting they try to use visual cues to compensate for semantic conflicts.
Key Results
- Conceptual Attribute Prediction: Accuracy drops notably when exceptional images are used with category-name queries (e.g., condition (1b) vs. (1a) in Figure 3i). For example, accuracy can fall from around 0.8 to 0.4 for some models.
- Instance Attribute Recognition: Performance also declines when exceptional images are presented without explicit exception cues (e.g., condition (2b) vs. (2a) in Figure 3ii). Accuracy can drop by 0.2 or more for some models.
- Pragmatic Violation is Stronger: The paper found that violating the pragmatic prior (incongruent image-text pairs) has a stronger negative impact on performance than violating the semantic prior (using an exception image for a general query).
- Shapley Values: V-SHAP (image contribution) increases for instance queries compared to concept queries, and is higher for exception images than typical images, indicating models recruit visual information more in these challenging scenarios.
Limitations and Future Work
- Limitations:
- The dataset's conceptual norms are limited to American English.
- The data collection focused on quality, potentially omitting rare or hard-to-photograph exceptions.
- The dataset primarily uses concrete object categories, and similar issues might arise with more abstract or social categories (e.g., stereotyping).
- Future Work: The authors suggest that their work is a step towards understanding VLM limitations and mitigating risks associated with overgeneralization. Further research could explore how to make VLM conceptual representations more robust to exceptions and how to better disentangle instance-level and conceptual-level reasoning.
Summary updated