We often think of images and words as distinct forms of communication, existing in separate realms of understanding. But what if they spoke a secret language, a shared vocabulary of concepts that underpins how artificial intelligence (AI) understands both? A groundbreaking study from researchers at Université Paris-Saclay, CEA, List reveals just that, offering a startling glimpse into the hidden connections between how AI processes visual and textual information.
Unveiling the Shared Vocabulary of AI
Clément Cornet, Romaric Besançon, and Hervé Le Borgne developed novel techniques to analyze the inner workings of AI models, using “sparse autoencoders” (SAEs). Think of SAEs as sophisticated linguistic detectives. They dissect the complex patterns of neural activity within AI models, identifying specific features that correspond to interpretable semantic concepts. Previous work using SAEs focused on comparing models within the same modality (e.g., comparing two image-processing AIs). This research goes much further, analyzing and comparing AIs trained on different modalities (images, text, and combinations of both).
The researchers’ innovation lies in two new tools: a weighted Max Pairwise Pearson Correlation (wMPPC) and a Comparative Sharedness measure. wMPPC assesses the similarity between concepts across different AI models, giving extra weight to those concepts used more frequently by the AI. The Comparative Sharedness measure goes even deeper, pinpointing specific concepts that one model shares more strongly with a certain class of models than others. For example, we can use these tools to pinpoint the concepts that a visual AI shares more strongly with language models than with other visual models.
A Deeper Look at Multimodal Models
The team applied these tools to a diverse group of 21 AI models, including large language models (LLMs) such as BERT and DeBERTa, visual foundation models (visual-only AIs) like DinoV2 and ViT, and multimodal models (AIs trained on both images and text) such as CLIP, DFN, and SigLIP2. The surprising result? The shared concepts between image and text AI were concentrated primarily in the final layers of each model. This suggests that the most meaningful semantic interpretation happens at the point where the AI has already integrated and processed its input.
The research also revealed differences in the alignment between image and text data across various datasets. The authors found that datasets with higher-quality image-text pairings (where the caption accurately describes the image) resulted in more overlap in the way different AIs understood the image and text. This implies that the quality of the training data is crucial in determining how much overlap in concepts different AIs will demonstrate.
The Impact of Text on Vision
Perhaps the most fascinating discovery was the identification of specific concepts unique to vision-language models (VLMs), which weren’t present in visual-only models. By using their Comparative Sharedness measure, the researchers were able to isolate these concepts. These included subtle yet meaningful groupings, such as:
- Age-related features: The VLM could differentiate images depicting children in various situations (birthday party, brushing teeth, playing baseball), associating each with a particular age group.
- Unusual pet behaviors: The VLM recognized and categorized images of pets engaged in uncommon activities (wearing hats, sitting on laptops) as distinct concepts, unlike visual-only models.
- Rooms of the house: The VLM formed clusters for various rooms (bedroom, bathroom, kitchen) based on visual features.
- Vehicles: The VLM connected distinct visual features of various train types (high-speed, freight, steam), indicating understanding of semantic similarity rather than just visual similarities.
- Geographical features: The model established connections between images representing a specific geographical region (e.g., different types of African animals or Italian foods).
- Concepts associated with actions: Remarkably, one feature clustered images of items associated with the verb “to ride” (horses, skis, bikes, surfboards) – demonstrating a deeper semantic connection that extends beyond pure visual characteristics.
Further investigation revealed that many of these VLMs’ unique visual concepts also showed strong correlations with features from language models trained on image captions. This suggests that the incorporation of text during the training of the VLMs fundamentally alters how these models understand visual information, going beyond simple image recognition to encompass higher-level semantic understanding. It’s as if the AI has learned to “read” into the images, gleaning deeper meaning through its textual training.
Implications and Future Directions
This work has profound implications for the field of Explainable AI (XAI), offering a deeper understanding of how multimodal AI models work. The ability to pinpoint and interpret the shared concepts between visual and textual representations promises to improve AI’s ability to translate information between modalities and to create more human-understandable explanations of AI decision-making processes. It opens doors to improved AI training by allowing researchers to better assess the image-text alignment quality of datasets, leading to improved models. The tools developed in this research could be utilized to identify and analyze concepts in a wider range of AI models, leading to more robust and nuanced XAI frameworks.
The researchers acknowledge limitations, including the focus on transformer-based models and the asymmetry of their wMPPC indicator. But this study provides an exciting stepping stone towards developing even more powerful methods for interpreting the intricate cognitive processes of AI. In essence, it’s a key step in deciphering the secret language spoken by the increasingly sophisticated minds we’re building.
