TLDR: This research introduces new methods, wMPPC and Comparative Sharedness, to quantitatively compare how visual, textual, and multimodal AI models share internal concepts. By training Sparse Autoencoders (SAEs) on 21 different encoders, the study found that cross-modal concept sharing primarily occurs in the models’ last layers and that text pretraining significantly influences visual models to learn high-level, text-related visual concepts.
In the rapidly evolving world of artificial intelligence, understanding how different models process and share information is crucial. A new research paper, “Explaining How Visual, Textual and Multimodal Encoders Share Concepts,” delves into this very challenge, offering novel ways to compare and interpret the internal workings of AI models across various data types.
Authored by Clément Cornet, Romaric Besançon, and Hervé Le Borgne from Université Paris-Saclay, CEA, List, this study introduces innovative tools to shed light on the complex “thoughts” of neural networks. Their work focuses on a technique called Sparse Autoencoders (SAEs), which are powerful for extracting human-understandable features from the hidden layers of these AI models.
Understanding Sparse Autoencoders (SAEs)
Imagine an AI model learning to recognize objects in images or understand sentences. As it processes information, it generates internal “activations” – complex numerical patterns that represent what it’s currently focusing on. SAEs act like a translator, taking these complex patterns and breaking them down into simpler, more interpretable “features” or “concepts.” For example, an SAE might identify a feature that consistently activates when the AI sees a “cat” or reads the word “happy.” The key here is “sparse,” meaning that for any given input, only a few of these features are active, making them easier to pinpoint and understand.
New Tools for Comparison: wMPPC and Comparative Sharedness
Previous research using SAEs often compared models only within the same type – for instance, two image recognition models. However, this new paper introduces two groundbreaking indicators that allow for comparisons across different modalities, such as visual models, text models, and even multimodal models that handle both images and text.
- Weighted Max Pairwise Pearson Correlation (wMPPC): This indicator measures the similarity between two different AI models based on the concepts extracted by SAEs. Unlike previous methods, wMPPC gives more weight to “important” features – those that are activated more frequently and with higher intensity. This helps in understanding which core concepts are truly shared between models.
- Comparative Sharedness: This tool goes a step further, allowing researchers to identify individual features within one model that are better shared with one class of models than another. For example, it can pinpoint a visual concept in an image model that is strongly linked to text understanding, but not to other purely visual models.
Key Findings from Extensive Experiments
The researchers applied these new tools in an extensive study involving 21 different AI encoders, including popular models like CLIP, DFN, SigLIP2 (multimodal), BERT, DeBERTa (text), and DinoV2, ViT, MambaVision (visual). They tested these models on various datasets, including general-purpose ones like COCO and Laion-2B, and a domain-specific one like Oxford-102 Flowers.
The study yielded several fascinating insights:
- Last Layers Hold the Key to Cross-Modal Sharing: The shared information between models of different modalities (e.g., image and text) is predominantly found in their final processing layers. This suggests that as models refine their understanding, they converge on more abstract, shared concepts.
- Dataset Quality Matters for Alignment: The wMPPC indicator revealed differences in how well image and text concepts align across different datasets. For instance, human-written captions in COCO led to better image-text alignment compared to internet-scraped captions from Laion-2B.
- Text Pretraining’s Profound Impact on Vision: Perhaps one of the most striking findings is the identification of visual concepts that are specific to Vision-Language Models (VLMs) like CLIP, but not present in traditional visual-only models. These “VLM-specific” visual features were found to be highly correlated with concepts learned by text encoders. This includes high-level semantic concepts such as “age-related situations,” “pets with unusual behaviors,” “rooms of the house,” “types of vehicles,” “old photos,” and even abstract concepts like “to ride” (which groups visually diverse objects like horses, skis, bikes, and jetskis). This strongly highlights the significant influence of text pretraining on how visual models understand and represent the world.
Also Read:
- Efficient Interpretability for Pruned LLMs Using Transferable Autoencoders
- How Data Characteristics Shape Concept Probing Performance
Implications for AI Understanding
This research provides a more systematic and quantitative approach to understanding the internal representations of large neural networks. By offering tools to compare concepts across different modalities, it opens new avenues for interpreting how AI models learn and generalize. The findings underscore the powerful impact of multimodal training, particularly text pretraining, on shaping the conceptual understanding of visual models. This work, detailed further in the paper available at arXiv:2507.18512, paves the way for more transparent and explainable AI systems.
