Skip to content

Explaining CLIP Zero-shot Predictions Through Concepts

🕒 Published (v1): 2026-03-30 09:31 UTC · Source: Arxiv · Venue: CVPR 2026 · link

Why this paper was selected

CVPR 2026; concept bottleneck explanations make CLIP zero-shot predictions transparent

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EZPC introduces a lightweight, trainable linear projection that maps CLIP's joint image-text embeddings into a shared human-interpretable concept space, enabling faithful concept-level explanations of zero-shot predictions without retraining CLIP or requiring concept supervision. Unlike prior concept bottleneck models (CBMs), EZPC preserves CLIP's open-vocabulary generalization while surfacing which specific visual/semantic concepts drive each classification decision. The method is trained with a matching loss (anchoring projection columns to known concept embeddings) and a reconstruction loss (preserving CLIP's similarity structure via KL divergence).

Problem

CLIP and similar VLMs achieve strong zero-shot generalization but produce opaque, entangled high-dimensional embeddings that offer no insight into why a prediction was made. Classical CBMs are interpretable but require dense concept supervision and operate in closed-world settings, failing to generalize to unseen classes. Recent CLIP-interpretability methods (e.g., SpLiCE, Z-CBM) either require expensive per-instance test-time optimization or large concept repositories with costly regression, and still lack a unified trainable decomposition that jointly preserves semantic fidelity and interpretability.

Method

EZPC learns a single projection matrix \(A \in \mathbb{R}^{d \times m}\) that maps CLIP's \(d\)-dimensional embeddings into an \(m\)-dimensional concept space, where each column of \(A\) corresponds to a human-interpretable concept direction (e.g., "has feathers", "made of metal"). Normalized image embeddings \(v_x\) and class text embeddings \(T\) are projected as \(c_x = v_x A\) and \(C_Y = T A\). Zero-shot classification is performed via dot product in concept space: \(\hat{y} = \arg\max_k \langle c_x, c_k \rangle\). Per-concept contribution scores are computed as the Hadamard product \(s_{x,k} = c_x \odot c_k\), making explanations faithful by construction (they directly compose the logit).

Training optimizes two objectives jointly: $\(\mathcal{L}_\text{total} = \mathcal{L}_\text{match} + \lambda \mathcal{L}_\text{recon}\)$ - Matching loss \(\mathcal{L}_\text{match}\): MSE between \(A\) and the matrix \(\Phi\) of CLIP text embeddings for the concept vocabulary, anchoring columns to interpretable directions. - Reconstruction loss \(\mathcal{L}_\text{recon}\): KL divergence between the concept-space similarity distribution \(\text{softmax}(c_i C_Y^\top)\) and CLIP's original similarity distribution \(\text{softmax}(v_i T^\top)\), enforcing semantic faithfulness.

Concept vocabularies are sourced from LF-CBM's GPT-3-generated attribute sets, ranging from 370 concepts (CUB) to 4,751 (ImageNet-1k).

Key Contributions

  • Trainable linear projection \(A\) that decomposes CLIP's joint vision-language space into a shared interpretable concept basis without additional supervision or retraining.
  • Dual training objective (matching + KL reconstruction) that simultaneously preserves interpretability and CLIP's predictive behavior.
  • Concept contribution scores \(s_{x,k} = c_x \odot c_k\) that are faithful by construction — they directly compose the classification logit.
  • Demonstrated cross-dataset transfer: projection trained on ImageNet-100 generalizes to CIFAR-100 and CUB without fine-tuning.
  • Spatial concept-region alignment analysis using patch-level features projected into concept space, with quantitative localization metrics (pointing accuracy, IoU).

Results

All results use CLIP RN50 backbone unless noted; generalized zero-shot harmonic mean (H) is the primary metric.

  • CIFAR-100: EZPC H = 0.403 vs. CLIP H = 0.408 (−0.005); Z-CBM H = 0.365; SpLiCE H = 0.270.
  • ImageNet-100: EZPC H = 0.682 vs. CLIP H = 0.693 (−0.011); Z-CBM H = 0.585; SpLiCE H = 0.389.
  • CUB-200-2011: EZPC H = 0.465 vs. CLIP H = 0.474 (−0.009); Z-CBM H = 0.189; SpLiCE H = 0.070.
  • ImageNet-1k: EZPC H = 0.481 vs. CLIP H = 0.530 (−0.049); Z-CBM H = 0.462; SpLiCE H = 0.300.
  • Places365: EZPC H = 0.352 vs. CLIP H = 0.362 (−0.010); Z-CBM H = 0.357; SpLiCE H = 0.282.
  • EZPC exceeds Z-CBM and SpLiCE by 10–15% harmonic mean on most datasets.
  • Backbone scaling: gap between CLIP and EZPC narrows with larger backbones (RN50 → ViT-B/32 → ViT-L/14 → SigLIP ViT-SO400M/14); SigLIP SO400M/14 shows EZPC H = 0.878 vs. CLIP H = 0.880.
  • Cross-dataset transfer (trained on ImageNet-100): generalized zero-shot H on CIFAR-100 is 0.409 (CLIP 0.380); on CUB is 0.515 (CLIP 0.526).
  • Concept localization (Indigo Bunting, positive concept "a blue-gray body"): 96.7% pointing accuracy; substantially higher IoU@10% and IoU@20% than negative concept "a red face".

Limitations

  • Performance gap on ImageNet-1k (~5% harmonic mean) suggests the concept decomposition scales less cleanly at large label-space size.
  • There is an inherent trade-off between quantitative fidelity (favored by large \(\lambda\)) and qualitative interpretability (favored by small \(\lambda\)); no single \(\lambda\) optimizes both simultaneously.
  • Some activated concepts reflect class-level statistical associations rather than visually present content in the specific image (e.g., "a hammock" for the class sea).
  • Concept vocabulary quality depends on the upstream GPT-3-generated sets from LF-CBM; noise in CLIP's embedding space can produce semantically inconsistent class-level explanations (e.g., "a pack of dholes" for Cardinal).
  • Evaluated primarily with CLIP RN50 for comparability with baselines; results with other backbones are provided but comparisons to Z-CBM/SpLiCE under those backbones are absent.
  • Linear projection may be insufficient to capture non-linear compositional concept interactions present in CLIP's representation.

Relevance to Vision-Language Models

EZPC directly targets the opacity of CLIP-style contrastive VLMs, which is a recognized bottleneck for deploying VLMs in high-stakes settings — the method provides a post-hoc yet training-time interpretability layer that preserves the model's zero-shot generalization. This work extends the line of CLIP interpretability research (alongside SpLiCE, Z-CBM, Gandelsman et al.) by offering a computationally efficient, training-time alternative to per-instance test-time optimization, making concept-level debugging of VLM predictions tractable at scale. The concept-region alignment analysis and cross-dataset transfer results are particularly relevant for researchers studying CLIP's internal semantic structure and grounding capabilities. For those tracking VLMs, EZPC establishes that faithfulness and open-vocabulary generalization need not be mutually exclusive, which matters for grounding, visual question answering, and trustworthy multimodal reasoning.