Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models

Technical Summary: Conceptual Embeddings: Sparse Disentanglement for Vision–Language Models

Problem Statement
Vision-language models (VLMs) such as CLIP, BLIP, and CoCa learn powerful high-dimensional shared embeddings, yet their internal semantic structure remains opaque. Existing interpretability methods face significant limitations:

• Local Explanations: Techniques like saliency maps or attribution methods identify relevant input regions but fail to reveal the global organization of the embedding space.
• Concept Alignment: Methods like TCAV or linear probing align embeddings with predefined concepts but often assume meaningful concepts are already linearly separable within the original geometry, limiting flexibility.
• Sparse Autoencoders (SAEs): While SAEs successfully extract interpretable, monosemantic features by learning overcomplete dictionaries, they expand the representation dimensionality. This expansion alters the original geometry of the embedding space, introduces redundancy, and produces lossy, non-invertible representations, making it difficult to faithfully relate explanations back to the underlying model.

The authors hypothesize that the apparent entanglement in VLM representations may stem from basis misalignment rather than insufficient dimensionality, suggesting that semantic structure could be revealed by reorganizing the space within its original dimensionality.

Methodology: CEDAR
The authors propose CEDAR (Conceptual Embedding Disentanglement via Adaptive Rotation), a post-hoc method that learns an invertible reparameterization of the embedding space without increasing dimensionality.

1. Invertible Transformation: CEDAR seeks an orthogonal linear operator $U$ (parameterized as $U = \exp(A - A^\top)$) that transforms the original embedding $z$ into a new space $\tilde{z} = U(z - b)$, where $b$ is the mean embedding. The orthogonality constraint ensures the transformation preserves the geometry and norm of the original space ($\|Uz\|_2 = \|z\|_2$) and is strictly invertible.
2. Sparsity Bottleneck: In the transformed space, a top-k operator ($\Pi_k$) is applied to retain only the $k$ coordinates with the largest absolute magnitudes, setting the rest to zero. This enforces a sparse representation where semantic information is concentrated in a small subset of dimensions.
3. Reconstruction Objective: The model is trained to minimize the reconstruction error between the original embedding and the sparse version mapped back to the original space:
   $$\min_U \mathbb{E}_z \left[ \| z - (U^{-1}\Pi_k(U(z - b)) + b) \|_1 \right]$$
   Because $U$ is invertible, the original embedding can be faithfully reconstructed, preserving the geometric structure essential for tasks like text-image alignment.
4. Training Strategy: To handle the non-smooth transition induced by the top-k operator, the authors employ a curriculum learning approach, starting with dense reconstructions and gradually transitioning to the target sparse regime. During training, $k$ is randomly sampled to improve robustness.
5. Interpretability Modes:
   • Concept-Based (CLIP-like): Individual axes in the transformed space are aligned with textual concepts using cosine similarity against a text encoder (e.g., CLIP text embeddings).
   • Free-Form (Generative): For generative models (e.g., BLIP), the sparse reconstructed embeddings are passed to the text decoder to generate natural language descriptions, allowing for sentence-level explanations derived from a small set of active coordinates.

Key Contributions

• Dimensionality-Preserving Disentanglement: Unlike SAEs, CEDAR achieves sparse, disentangled representations without expanding the embedding dimension, thereby avoiding geometric distortion and redundancy.
• Invertibility: The method guarantees faithful reconstruction of the original embedding, ensuring explanations remain directly linked to the original model's representation space.
• Unified Framework: CEDAR supports both concept-based alignment (matching neurons to text) and free-form generation (decoding sparse vectors to text) within a single framework.
• Hypothesis Validation: The work provides empirical evidence that semantic structure in VLMs can be uncovered through a suitable change of basis, challenging the necessity of overcomplete expansions for interpretability.

Results
The authors evaluate CEDAR against Sparse Autoencoder variants (MSAE, TopK, ReLU, BatchTopK) and dense baselines on standard benchmarks (ImageNet-1K) and through user studies.

• Reconstruction-Sparsity Trade-off: CEDAR achieves a competitive trade-off between reconstruction error (FVU) and sparsity (number of active features $K$). In low-dimensional regimes, it often requires fewer active components to achieve comparable reconstruction fidelity to other methods.
• Semantic Preservation: Metrics such as Centered Kernel Nearest Neighbor Alignment (CKNNA) and Linear Probing cross-entropy (LP) indicate that CEDAR preserves semantic structure and downstream task performance comparable to the strongest baselines (e.g., BatchTopK) and the dense model.
• User Study:
  • Preference: Participants consistently preferred CEDAR's concept explanations over MSAE, finding them more representative of image content.
  • Relevance: CEDAR's suggested concepts were selected as relevant by human evaluators significantly more often than those from MSAE.
  • Description Quality: When used to generate captions via a text decoder, CEDAR's sparse representations produced descriptions that "describe the image" at rates comparable to the full dense model, vastly outperforming sparse models without disentanglement.
• Qualitative Analysis: Visual examples show CEDAR identifying coherent, high-level semantic concepts (e.g., "dog," "husky," "wooden") and generating accurate captions, whereas baselines often retrieved noisy or lower-level features.

Significance and Claims
The paper claims that CEDAR demonstrates that the apparent entanglement in modern vision-language embeddings is largely a result of basis misalignment rather than a lack of dimensionality. By learning a global, invertible coordinate system, the method reveals a compositional structure where semantic factors are concentrated in a small number of axis-aligned dimensions.

The authors position CEDAR as a flexible, model-agnostic framework that unifies concept-based and free-form interpretability. They emphasize that their approach eliminates the need for overcomplete expansions, offering a more faithful and information-efficient way to understand and manipulate learned representations. The work suggests that suitable reparameterization can resolve entanglement while preserving the geometric integrity of the original model.