---
title: Explain Any Concept (EAC) in Computer Vision
url: https://www.emergentmind.com/topics/explain-any-concept-eac
type: topic
---

# Explain Any Concept (EAC) in Computer Vision

Explain Any Concept (EAC) is a local, concept-based explainable AI method for computer vision that explains a model’s prediction on a single input image using human-understandable image concepts rather than raw pixels. In its original formulation, EAC combines automatic image segmentation from the Segment Anything Model (SAM), Shapley-value attribution over segmented concepts, and a lightweight per-input equivalent (PIE) surrogate model to make concept-level explanation computationally tractable. In that original sense, “any concept” means any visual concept that can be segmented from the input image, not an arbitrary user-defined textual concept [2305.10289].

## 1. Concept-based explanation before EAC

EAC emerged from a line of work that criticized pixel-level explanation as too local and too low-level for stable semantic interpretation. ACE, or “Automated Concept-based Explanations,” formulated three desiderata for concept explanations—meaningfulness, coherency, and importance—and operationalized them by segmenting class images at multiple resolutions, clustering segments in representation space, removing outliers, and scoring discovered concepts with TCAV-style importance measures. In ACE, concepts are dataset-level recurring visual patterns rather than per-image regions, and the method is explicitly global rather than local [1902.03129].

The original EAC paper adopted the same general premise that concepts are more aligned with human reasoning than pixels, but changed the unit of explanation. Instead of discovering recurring concepts across many images by clustering superpixels, EAC constructs a concept set separately for each image by running SAM and treating the resulting segments as candidate concepts. This shift matters because it trades dataset-level concept discovery for input-specific semantically meaningful regions, and thereby targets explanation of a single prediction rather than explanation of a class in the aggregate [2305.10289].

## 2. Architecture of the original EAC method

The original EAC pipeline has three phases. In phase one, SAM is applied to an input image \(x\) to produce an image-specific concept set
\[
\mathcal{C}=\{c_1,c_2,\ldots,c_n\}.
\]
Each \(c_i\) is a segmented visual instance or region, and these segments are treated as the explanatory vocabulary for that image. This is the operational meaning of “Explain Any Concept” in the 2023 method: any SAM segment in the image can become an explanation concept [2305.10289].

Phase two introduces the PIE surrogate. Because Shapley-value attribution over many concept coalitions would be expensive if every masked variant had to be evaluated by the original target model \(f\), EAC trains a small surrogate \(f'\) only for the current image and its concept perturbations. The surrogate is written as
\[
f'(b) := f_{FC}(\mathbf{h}(b)),
\]
where \(b \in \{0,1\}^n\) is a binary concept-indicator vector, \(\mathbf{h}\) is the trainable surrogate feature extractor, and \(f_{FC}\) is the frozen final fully connected layer of the original target network. Training data are generated by sampling concept subsets, masking the image accordingly, querying the target model on the masked image, and optimizing \(\mathbf{h}\) with cross-entropy while keeping \(f_{FC}\) fixed [2405.11837].

Phase three performs concept attribution with Shapley values. For a coalition \(S \subseteq \mathcal{C}\), the utility is defined by masking out concepts not in \(S\):
\[
u(S)\coloneqq f(\text{mask}(x, \mathcal{C}\setminus S)).
\]
For concept \(c_i\), the marginal contribution to coalition \(S\) is
\[
\Delta_{c_i}(S) = u(S\cup \{c_i\}) - u(S),
\]
and the practical estimator is the Monte Carlo approximation
\[
\hat{\phi}_{c_i}(x) = \frac{1}{K} \sum_{k=1}^K \Delta_{c_i}(S_k).
\]
The final explanation is a subset \(E \subseteq \mathcal{C}\) formed from the highest-contributing concepts, and the output is visualized by masking out concepts in \(\mathcal{C}\setminus E\). The method is therefore local, prediction-specific, and concept-based rather than class-global [2305.10289].

## 3. Empirical behavior, faithfulness, and surrogate design

On ImageNet and COCO with ResNet-50, the original EAC outperformed the reported superpixel-based baselines on insertion and deletion AUC. The reported insertion AUCs were \(83.400\) on ImageNet and \(83.404\) on COCO; the corresponding deletion AUCs were \(23.799\) and \(16.640\). The paper highlights improvements over the second-best DeepLIFT baseline of \(8.165\%\) on ImageNet insertion and \(5.205\%\) on COCO insertion. A human evaluation over 200 images reported consensus on 184 images, with EAC favored on 137 images, or \(74.5\%\), which the paper presents as evidence of improved understandability [2305.10289].

The computational bottleneck is the Shapley loop. Direct original-model Monte Carlo Shapley computation took more than 24 hours per image, which made the unapproximated version impractical. The PIE ablation showed that parameter sharing with the target model’s classifier head was crucial: on ImageNet, PIE obtained insertion \(81.78\) and deletion \(12.47\) in about 245 seconds, whereas PIE without parameter sharing obtained insertion \(50.40\) and deletion \(32.87\); the linear surrogate was faster but less faithful, with insertion \(78.11\) and deletion \(14.08\) [2305.10289].

A later refinement, “Improving the Explain-Any-Concept by Introducing Nonlinearity to the Trainable Surrogate Model,” replaced EAC’s single trainable linear layer with two trainable fully connected layers separated by a nonlinearity. The tested activations were \(\tanh\), sigmoid, and ReLU, with \(\tanh\) performing best. On ImageNet, the \(\tanh\) variant reported insertion AUC \(85.60\) and deletion AUC \(22.10\), compared with original EAC’s \(84.04\) and \(23.56\); on COCO, it reported \(85.22\) and \(16.11\), compared with \(83.87\) and \(17.31\). The runtime increase was small, from 250 to 255 seconds on ImageNet and from 212 to 217 seconds on COCO. In this refinement, “improves performance” specifically means more faithful concept ranking under insertion and deletion evaluation, not a change in concept discovery itself [2405.11837].

## 4. From segmented image concepts to open-vocabulary concept spaces

The phrase “Explain Any Concept” subsequently became associated with a broader effort to relax the concept bottlenecks that remain in the original EAC formulation. One direction is automatic concept acquisition without manual concept labels. ACEM, or “Automatic Concept Embedding Model,” uses an ACE-inspired discovery stage to segment images with SLIC, cluster segments with k-means, represent each discovered concept as a hypersphere in activation space, pseudo-label the training set, and then train a standard CEM on the resulting \(X \times C \times Y\) data. ACEM therefore removes the need for concept annotations during training, although it does not support arbitrary user-defined concepts at test time. On MNIST-ADD, ACEM with pre-trained activation space reported task accuracy \(0.947 \pm 0.001\) and CAS \(0.96\), close to supervised CEM’s \(0.953 \pm 0.002\) and \(0.97\); on CUB, ACEM reported \(0.76\) task accuracy and \(0.83\) CAS, with degradation attributed to irrelevant background concepts such as water or forest [2309.03970].

A second direction is language-guided concept training. LG-CAV uses a pretrained vision-language model such as CLIP to generate soft concept activations on a shared probe set, then trains a concept activation vector \(v_c\) inside the target model so that
\[
\mathcal{L}_{\rm LG\text{-}CAV} =
\frac{1}{\|\mathcal{R}\|}\sum_{x\in\mathcal{R}}
\big({\rm Act}_{v_c}(f(x)) - {\rm Act}_{g_{\rm text}(c)}(g_{\rm img}(x))\big)^2.
\]
The method adds Gaussian Alignment, Concept Ensemble, and Deviation Sample Reweighting, and it does not require concept-labeled images in its pure language-guided mode. On Broden, the full LG-CAV variant substantially improved both concept accuracy and concept-to-class accuracy over original CAV, Text-to-Concept, and OA-TCAV; for example, on ResNet-18 it reported concept accuracy \(77.45\) and concept-to-class accuracy \(24.58\), versus \(68.92\) and \(6.20\) for original CAV. This extends EAC-like aspirations toward open-vocabulary concept queries, although each new concept still requires a separate optimization [2410.10308].

A third direction is explicit open-vocabulary concept bottlenecks after training. OpenCBM aligns a trainable image feature extractor to CLIP’s image embedding space with prototype-based alignment, then reconstructs each learned class weight vector \(\mathbf{v}_c\) as
\[
\mathbf{v}_c \approx \sum_{i=1}^{k}\alpha_{ci}\mathbf{t}_i,
\]
where \(\mathbf{t}_i\) are user-provided text concept embeddings. Because the concept basis is supplied after training, the method supports adding, removing, or replacing concepts post hoc, and it introduces residual-based search for missing concepts when user concepts do not span the classifier well. On CUB-200-2011, OpenCBM reported \(83.3\%\) accuracy, compared with \(74.3\%\) for Label-free CBM and \(79.9\%\) for the black-box ResNet18 baseline, and the paper describes it as the first CBM with concepts of open vocabularies [2408.02265].

## 5. Related extensions beyond standard 2D classification

Although the original EAC paper is centered on 2D vision classification with SAM-generated segments, the broader concept-based explanation literature has already moved into more difficult regimes such as multi-modal volumetric detection and classification. A representative example is the PSMA-PET/CT study on metastatic prostate cancer, which used regression concept activation on 3D PET/CT volumes. There the concepts were continuous radiomic quantities, such as lesion volume, 3D circularity, PET intensity, and texture statistics, rather than segmented image parts. The reported combined system achieved sensitivity \(80\%\) at \(1.78\) false positive per patient; global explanations indicated that detection focused on CT for anatomical location and on PET for confidence, while local explanations helped distinguish true positives from false positives. This line of work shows that concept-based explanation is not tied to 2D pixel saliency or to per-image instance masks, and it highlights a parallel expansion from feature attribution toward semantically meaningful latent factors in high-stakes medical imaging [2208.02555].

Within that wider landscape, EAC occupies a specific design point. Its concepts are local image segments rather than latent directions, its explanation target is a single prediction rather than an entire dataset, and its computational novelty lies in the PIE surrogate rather than in concept learning from labels or language. The relation to neighboring approaches is therefore structural: EAC supplies an efficient local concept-attribution pipeline for segmented images, whereas methods such as LG-CAV and OpenCBM address concept acquisition and open-vocabulary access at the representation level [2305.10289].

## 6. Limitations, ambiguities, and open problems

Several limitations are explicit in the original EAC formulation. The quality of the explanation depends on SAM; if SAM produces poor or ill-identified concepts, EAC degrades accordingly. The paper is mainly evaluated on image classification, not on broader visual tasks, and it leaves several concept-structure issues unspecified, including overlap handling among SAM masks, granularity selection, merging or splitting of segments, background handling, mask fill policy, and the exact architecture of the trainable mapping \(h\). Even with PIE, the reported runtimes remain on the order of hundreds of seconds per image, and the method still relies on Monte Carlo approximation of Shapley values rather than exact coalition enumeration [2305.10289].

Later extensions resolve some bottlenecks while exposing others. The nonlinear-surrogate refinement improves insertion and deletion faithfulness but is presented as preliminary, explores only one additional hidden layer with three activation functions, and does not report human-subject evaluation of understandability [2405.11837]. ACEM removes manual concept labels, but its CUB results show that automatically discovered concepts can be dominated by nuisance structure such as background water or forest, so human-meaningful semantics are not guaranteed [2309.03970]. LG-CAV reduces dependence on concept-labeled images, yet it inherits the semantic coverage and prompt sensitivity of the underlying vision-language model, and each concept still requires per-concept training [2410.10308]. OpenCBM supports adding, removing, and replacing concepts after training, but it depends on the quality of CLIP’s image-text space, and its residual search can recover concepts such as “melodious song,” which are semantically meaningful but not directly visible in the image [2408.02265].

Taken together, these results suggest that EAC is best understood not as a single closed method but as a family of design commitments within concept-based XAI: replace pixel heatmaps with semantically meaningful explanatory units, expose model behavior through concept-level manipulations, and make those explanations faithful enough to survive perturbation-based evaluation. The unresolved question is what the concept vocabulary should be—per-image segments, automatically discovered clusters, language-guided directions, or open-vocabulary text concepts—and how that vocabulary should be grounded so that faithfulness, human semantics, efficiency, and post-deployment flexibility can coexist.

Source: https://www.emergentmind.com/topics/explain-any-concept-eac