Skip to content

Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

FIxLIP is a game-theoretic framework that explains cosine similarity predictions in vision-language encoders (CLIP, SigLIP) by computing weighted Banzhaf interactions between all pairs of image patches and text tokens, capturing cross-modal and intra-modal relationships that first-order saliency maps miss. A cross-modal sampling strategy reduces model inference calls by 5–20× over Shapley-based approaches. Experiments on MS COCO and ImageNet-1k show FIxLIP consistently outperforms first-order attribution baselines across three purpose-built evaluation metrics.

Problem

Popular explanation methods for CLIP-style encoders (gradient, attention, and information-based saliency maps) assign importance only to individual tokens (first-order attributions), ignoring pairwise cross-modal interactions between image patches and text tokens. This single-modal view is unfaithful to the similarity function and cannot disentangle how distinct visual regions correspond to distinct words in multi-object scenes. Additionally, naive token masking produces out-of-distribution inputs (unrecognizable images, ambiguous captions) that distort game values.

Method

FIxLIP frames similarity explanation as a cooperative game: given input pair (x_I, x_T) and a baseline (b_I, b_T), the game value ν(M) is the cosine similarity of the masked image and text. An explanation e decomposes ν into a constant, per-token attributions e_i, and pairwise interaction terms e_{i,j} via a 2-additive approximation, minimizing p-faithfulness F_p—a Bernoulli(p)-weighted squared error over all masks that upweights in-distribution (few tokens masked) inputs when p > 0.5. This yields FIxLIP-p as the weighted Banzhaf interaction index, estimated via weighted least squares (WLS) over Monte Carlo–sampled masks.

To scale computation, the cross-modal estimator independently samples m_I image masks and m_T text masks, then evaluates all m_I × m_T combinations using vectorized cosine similarity after independent encoder passes (exploiting f = cos(f_I(x_I), f_T(x_T))). This turns m sequential model calls into m_I + m_T calls yielding m_I·m_T game values. Large-scale deployment uses two-step filtering: first-order top-k tokens are selected, then second-order interactions among them are computed.

Three new metrics are introduced: (1) p-faithfulness correlation (Spearman rank correlation between ν and ν̂ over sampled masks), (2) Area between Insertion/Deletion curves (AID, generalised to subset-optimal selection under the 2-additive model), and (3) Pointing Game Recognition (PGR, ratio of correctly signed cross-modal interactions across multi-object scenes).

Key Contributions

  • FIxLIP-p: first application of weighted Banzhaf interactions to vision-language encoder explainability; p controls in-distribution bias during explanation
  • Cross-modal estimator: theoretically unbiased, variance-bounded estimator achieving 20× speedup in model inferences and 5× end-to-end speedup over model-agnostic Shapley estimator
  • Three second-order evaluation metrics: p-faithfulness correlation, AID, and PGR, all extended from first-order to interaction-level explanations
  • Empirical finding: second-order explanations are necessary—first-order methods fail PGR sanity checks on multi-object inputs, while FIxLIP does not
  • Model comparison utility: FIxLIP reveals SigLIP-2 is more faithfully explainable via cross-modal interactions than CLIP of the same architecture

Results

  • AID on MS COCO (CLIP ViT-B/32): FIxLIP outperforms all baselines (GAME, Grad-ECLIP, exCLIP) in recovering optimal token subsets under insertion/deletion; Grad-ECLIP and exCLIP fail to capture nonlinear importance rankings
  • PGR on ImageNet-1k (CLIP ViT-B/32):
  • FIxLIP (Shapley interactions): .83 / .82 / .84 / .86 for 1–4 objects vs. Grad-ECLIP .68 / .45 / .33 / .28
  • exCLIP scores slightly higher on multi-object PGR (.88–.92) due to cross-modal specialisation, but omits intra-modal terms, degrading AID and faithfulness
  • p-faithfulness correlation: FIxLIP reaches near-perfect R² ≈ 1.0; first-order baselines (Grad-ECLIP, GAME) plateau at µ_corr ≈ 0.5
  • Computational speedup: cross-modal estimator achieves 20× speedup in game evaluations, 5× in full pipeline vs. model-agnostic estimator
  • Model comparison: SigLIP-2 (ViT-B/32) PGR .90 vs. CLIP .83 for 1 object; pattern holds across ViT-B/16; larger models (ViT-L/16) are harder to explain faithfully

Limitations

  • Faithfulness degrades as the number of input tokens grows (larger image resolutions); ViT-B/16 results are weaker than ViT-B/32
  • Subset selection for large token counts uses a greedy heuristic; near-optimal selection algorithms remain unexplored
  • No human evaluation of visual usability (user studies deferred to future work)
  • Restricted to second-order interactions; efficiently extending to higher-order or tri-modal (image–text–audio) games is an open problem

Relevance to Vision-Language Models

FIxLIP directly addresses the interpretability deficit of CLIP-style encoders, which serve as the visual backbone of virtually all modern VLMs (Flamingo, LLaVA, GPT-4V). Understanding why a VLE assigns high similarity to an image-text pair—including known failure modes like CLIP-blind pairs—is prerequisite to debugging VLMs in medical imaging and safety-critical settings. The weighted Banzhaf interaction framework is model-agnostic and has been validated on CLIP and SigLIP/SigLIP-2, covering the dominant pre-training paradigms. The PGR-based model comparison (CLIP vs. SigLIP-2) introduces a new axis for evaluating encoder quality beyond downstream task accuracy, relevant to architecture selection for VLM development.