Interpreting the Linear Structure of Vision-language Model Embedding Spaces

• URL: http://arxiv.org/abs/2504.11695v1
• Date: Wed, 16 Apr 2025 01:40:06 GMT
• Title: Interpreting the Linear Structure of Vision-language Model Embedding Spaces
• Authors: Isabel Papadimitriou, Huangyuan Su, Thomas Fel, Naomi Saphra, Sham Kakade, Stephanie Gil,
• Abstract summary: We train and release sparse autoencoders (SAEs) on the embedding spaces of four vision-language models.<n>We find that SAEs are better at reconstructing the real embeddings, while also able to retain the most sparsity.<n>We also show that the key commonly-activating concepts extracted by SAEs are remarkably stable across runs.
• Score: 15.065914257384204
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Vision-language models encode images and text in a joint space, minimizing the distance between corresponding image and text pairs. How are language and images organized in this joint space, and how do the models encode meaning and modality? To investigate this, we train and release sparse autoencoders (SAEs) on the embedding spaces of four vision-language models (CLIP, SigLIP, SigLIP2, and AIMv2). SAEs approximate model embeddings as sparse linear combinations of learned directions, or "concepts". We find that, compared to other methods of linear feature learning, SAEs are better at reconstructing the real embeddings, while also able to retain the most sparsity. Retraining SAEs with different seeds or different data diet leads to two findings: the rare, specific concepts captured by the SAEs are liable to change drastically, but we also show that the key commonly-activating concepts extracted by SAEs are remarkably stable across runs. Interestingly, while most concepts are strongly unimodal in activation, we find they are not merely encoding modality per se. Many lie close to - but not entirely within - the subspace defining modality, suggesting that they encode cross-modal semantics despite their unimodal usage. To quantify this bridging behavior, we introduce the Bridge Score, a metric that identifies concept pairs which are both co-activated across aligned image-text inputs and geometrically aligned in the shared space. This reveals that even unimodal concepts can collaborate to support cross-modal integration. We release interactive demos of the SAEs for all models, allowing researchers to explore the organization of the concept spaces. Overall, our findings uncover a sparse linear structure within VLM embedding spaces that is shaped by modality, yet stitched together through latent bridges-offering new insight into how multimodal meaning is constructed.

Related papers

• SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability [9.90112908284836]
  We introduce (Sparse Autoencoders for Aligned Representation of Concepts) a new framework that learns a single, unified latent space shared across diverse architectures.<n>On Open Images, dramatically improves concept alignment, achieving a Jaccard similarity of 0.80, more than tripling the alignment compared to previous methods.
  arXiv  Detail & Related papers  (2025-07-07T22:29:00Z)
• Cross-Modal and Uncertainty-Aware Agglomeration for Open-Vocabulary 3D Scene Understanding [58.38294408121273]
  We propose Cross-modal and Uncertainty-aware Agglomeration for Open-vocabulary 3D Scene Understanding dubbed CUA-O3D.<n>Our method addresses two key challenges: (1) incorporating semantic priors from VLMs alongside the geometric knowledge of spatially-aware vision foundation models, and (2) using a novel deterministic uncertainty estimation to capture model-specific uncertainties.
  arXiv  Detail & Related papers  (2025-03-20T20:58:48Z)
• Exploring the Small World of Word Embeddings: A Comparative Study on Conceptual Spaces from LLMs of Different Scales [47.52062992606549]
  A conceptual space represents concepts as nodes and semantic relatedness as edges.<n>We construct a conceptual space using word embeddings from large language models of varying scales.<n>We analyze conceptual pairs, WordNet relations, and a cross-lingual semantic network for qualitative words.
  arXiv  Detail & Related papers  (2025-02-17T02:52:07Z)
• Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment [6.614005142754584]
  Universal Sparse Autoencoders (USAEs) are a framework for uncovering and aligning interpretable concepts spanning multiple deep neural networks.<n>USAEs learn a universal concept space that can reconstruct and interpret the internal activations of multiple models at once.
  arXiv  Detail & Related papers  (2025-02-06T02:06:16Z)
• GSSF: Generalized Structural Sparse Function for Deep Cross-modal Metric Learning [51.677086019209554]
  We propose a Generalized Structural Sparse to capture powerful relationships across modalities for pair-wise similarity learning. The distance metric delicately encapsulates two formats of diagonal and block-diagonal terms. Experiments on cross-modal and two extra uni-modal retrieval tasks have validated its superiority and flexibility.
  arXiv  Detail & Related papers  (2024-10-20T03:45:50Z)
• Disentangling Dense Embeddings with Sparse Autoencoders [0.0]
  Sparse autoencoders (SAEs) have shown promise in extracting interpretable features from complex neural networks. We present one of the first applications of SAEs to dense text embeddings from large language models. We show that the resulting sparse representations maintain semantic fidelity while offering interpretability.
  arXiv  Detail & Related papers  (2024-08-01T15:46:22Z)
• SSPA: Split-and-Synthesize Prompting with Gated Alignments for Multi-Label Image Recognition [71.90536979421093]
  We propose a Split-and-Synthesize Prompting with Gated Alignments (SSPA) framework to amplify the potential of Vision-Language Models (VLMs) We develop an in-context learning approach to associate the inherent knowledge from LLMs. Then we propose a novel Split-and-Synthesize Prompting (SSP) strategy to first model the generic knowledge and downstream label semantics individually.
  arXiv  Detail & Related papers  (2024-07-30T15:58:25Z)
• Deep Boosting Learning: A Brand-new Cooperative Approach for Image-Text Matching [53.05954114863596]
  We propose a brand-new Deep Boosting Learning (DBL) algorithm for image-text matching. An anchor branch is first trained to provide insights into the data properties. A target branch is concurrently tasked with more adaptive margin constraints to further enlarge the relative distance between matched and unmatched samples.
  arXiv  Detail & Related papers  (2024-04-28T08:44:28Z)
• Identifying Interpretable Subspaces in Image Representations [54.821222487956355]
  We propose a framework to explain features of image representations using Contrasting Concepts (FALCON) For a target feature, FALCON captions its highly activating cropped images using a large captioning dataset and a pre-trained vision-language model like CLIP. Each word among the captions is scored and ranked leading to a small number of shared, human-understandable concepts.
  arXiv  Detail & Related papers  (2023-07-20T00:02:24Z)
• ONE-PEACE: Exploring One General Representation Model Toward Unlimited Modalities [71.15303690248021]
  We release ONE-PEACE, a highly model with 4B parameters that can seamlessly align and integrate representations across vision, audio, and language modalities. The architecture of ONE-PEACE comprises modality adapters, shared self-attention layers, and modality FFNs. With the scaling-friendly architecture and pretraining tasks, ONE-PEACE has the potential to expand to unlimited modalities.
  arXiv  Detail & Related papers  (2023-05-18T17:59:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.