CAVLI: Concept-Based Local Interpretability

• CAVLI is a local concept-based interpretability framework that combines TCAV and LIME to quantify how much a classifier’s decision relies on human-interpretable image regions via the Concept Dependency Score.
• It employs counterfactual reasoning through superpixel masking to generate perturbed images, enabling the creation of both concept and decision heatmaps for local analysis.
• Experimental evaluations on ImageNet and CelebA show that CAVLI effectively distinguishes meaningful semantic features from spurious cues, making it valuable for model debugging and bias auditing.

Searching arXiv for papers mentioning CAVLI and closely related work to ground the article. CAVLI is a local concept-based interpretability framework for image classifiers introduced in Chapter 3 of the thesis "Understanding and evaluating computer vision models through the lens of counterfactuals" (Shukla, 28 Aug 2025). It combines attribution-based explanations provided by LIME with concept-level representations provided by TCAV in order to quantify how strongly a classifier’s decision depends on a human-interpretable concept. The method is explicitly framed through counterfactual reasoning: it masks superpixels, observes changes in concept sensitivity and model output, and measures whether the regions associated with a concept overlap with the regions that support the prediction. Its central quantitative output is the Concept Dependency Score, or CDS, which is intended to distinguish reliance on meaningful semantic structure from reliance on spurious cues such as background context or protected attributes (Shukla, 28 Aug 2025).

1. Definition, motivation, and conceptual scope

CAVLI addresses a specific interpretability problem in computer vision: when a classifier predicts a class label for a particular image, one may wish to know whether the prediction depends on a meaningful semantic concept or on an irrelevant cue. The motivating example is a classifier predicting "cow": the relevant question is whether the model is using animal features such as body shape and texture, or instead relying on a spurious concept such as grassland in the background. The same logic applies to socially sensitive concepts in bias auditing, such as gender or race in facial or activity recognition settings (Shukla, 28 Aug 2025).

The framework is motivated by two limitations in prior explanation families. Standard pixel-level saliency methods do not directly answer whether a decision depends on a human-defined concept, while global concept methods such as TCAV can indicate whether a class is sensitive to a concept on average but not how that concept matters for a specific image and prediction. CAVLI therefore occupies a local, per-image regime: there is a trained image classifier, a concept is defined using example images, and the objective is to determine whether the model relied on that concept in the specific decision under inspection (Shukla, 28 Aug 2025).

The thesis treats this as a counterfactual question. CAVLI does not generate semantic edits such as a cow without grass; instead, it creates counterfactual-like perturbations by masking superpixels. Each perturbation asks a local question of the form: what would the model predict if this region were absent? This operationalization makes the method post hoc, image-specific, and concept-aware, while remaining compatible with ordinary trained classifiers rather than requiring retraining or generative intervention (Shukla, 28 Aug 2025).

The framework rests on three assumptions stated in the thesis. First, a concept can be represented in internal activation space using TCAV. Second, image regions can be localized meaningfully using superpixels. Third, stronger spatial overlap between concept-relevant regions and decision-relevant regions indicates stronger dependence of the decision on the concept. The thesis later notes that this remains an associational notion rather than full causal dependence, which is a central qualification in any technical interpretation of CAVLI (Shukla, 28 Aug 2025).

2. Hybrid TCAV–LIME construction

CAVLI is a hybrid TCAV-LIME pipeline built around two localized maps over the same superpixel decomposition of an image: a concept heatmap and a decision heatmap. The chapter defines a trained neural network $F$ on a dataset $X$ with labels $Y$, and denotes the activation at layer $l$ by $h_l(f_l(x_i))$. CAVLI first uses TCAV to represent a human-defined concept in activation space by learning a Concept Activation Vector $v_c$, and then computes conceptual sensitivity as a directional derivative along that concept direction: $CS^l_C(F, x_i) = \nabla h_l(f_l(x_i))^T v_c .$ The TCAV score is then the fraction of inputs for a class with positive conceptual sensitivity (Shukla, 28 Aug 2025).

Localization is introduced through superpixels. The image is partitioned into $r$ superpixels using SLIC, and CAVLI generates perturbed images by masking random subsets of those regions. These perturbations are explicitly interpreted as counterfactual-like variants of the input. For each perturbed image $x_{ij}$, the method computes a conceptual sensitivity value

$z_{ij} = CS^l_C(F, x_{ij}) .$

A local surrogate model is then fit over superpixel presence to predict these $X$0 values. The resulting coefficients correspond to superpixel-level concept association, and their visualization is the concept heatmap (Shukla, 28 Aug 2025).

In parallel, CAVLI runs standard LIME on classifier outputs for the same perturbed images. For each perturbation, the classifier is queried as

$X$1

A second local surrogate model is fit to predict these outputs from superpixel presence. Its coefficients represent the contribution of each superpixel to the decision and are visualized as the decision heatmap. CAVLI then compares the two coefficient vectors by Pearson correlation: $X$2 A larger $X$3 indicates greater overlap between the regions associated with the concept and the regions driving the prediction (Shukla, 28 Aug 2025).

The final score is the Concept Dependency Score: $X$4 In the intended interpretation, $X$5 captures local overlap between concept-relevant and decision-relevant regions, while $X$6 captures how strongly the concept direction matters in the representation. Their product is therefore high when the concept is both representationally salient and spatially aligned with the decision. High positive CDS indicates strong dependence of the prediction on the concept, while low or negative CDS indicates little dependence or anti-alignment (Shukla, 28 Aug 2025).

The thesis provides the method as a nine-step procedure: train a TCAV model for a concept $X$7, decompose the image into homogeneous superpixels, create perturbed images by masking random subsets, compute conceptual sensitivities for those perturbations, fit a local surrogate to produce the concept heatmap, query the classifier on the same perturbations, fit a second surrogate to produce the decision heatmap, compute Pearson correlation between the two coefficient vectors, and finally compute the CDS. This sequence is the complete algorithmic specification given in the source (Shukla, 28 Aug 2025).

3. Experimental setting and empirical behavior

CAVLI is evaluated on ImageNet and CelebA. For ImageNet, the thesis uses GoogleNet, ResNet-50, and InceptionNet, follows settings similar to Kim et al. and Schrouff et al., focuses on the classes Zebra and Basketball, uses 100 images per set, and conducts all experiments on the penultimate layer. The main quantitative baseline is TCAV itself: the chapter compares mean CDS against global TCAV score, with qualitative heatmaps used for local inspection (Shukla, 28 Aug 2025).

On ImageNet/Zebra, the reported concepts are stripes, grassland, indoor, and horse. The reported values are as follows. For GoogleNet, mean CDS and TCAV are 0.17 and 0.78 for stripes, 0.26 and 0.62 for grassland, 0.13 and 0.12 for indoor, and 0.02 and 0.41 for horse. For ResNet, they are 0.23 and 0.87 for stripes, 0.26 and 0.81 for grassland, 0.11 and 0.48 for indoor, and 0.11 and 0.51 for horse. For InceptionNet, they are 0.45 and 0.84 for stripes, 0.21 and 0.71 for grassland, $X$8 and 0.43 for indoor, and 0.16 and 0.35 for horse. The thesis interprets these results as showing that mean CDS broadly resembles TCAV trends while remaining locally meaningful, and notes in particular that ResNet and GoogleNet show highest mean CDS for stripes and lowest mean CDS for indoor, while InceptionNet shows grassland more strongly associated than expected (Shukla, 28 Aug 2025).

On ImageNet/Basketball, the analyzed concepts include ball, jersey, female, and race; the thesis notes that the race concept classifier was trained with positive class images of African American faces. The reported values are 0.29 and 0.56 for ball, 0.38 and 0.93 for jersey, $X$9 and 0.26 for female, and 0.24 and 0.46 for race in GoogleNet; 0.27 and 0.68 for ball, 0.21 and 0.46 for jersey, $Y$0 and 0.45 for female, and 0.22 and 0.73 for race in ResNet; and 0.41 and 0.87 for ball, 0.05 and 0.31 for jersey, 0.09 and 0.31 for female, and 0.18 and 0.57 for race in InceptionNet. The chapter states that these results further confirm previous findings of a correlation between decisions on the basketball class and the concept race (Shukla, 28 Aug 2025).

The principal qualitative demonstration is the cow–grassland example. Figure 1.3 presents the original image, a concept heatmap for grassland, and a decision heatmap for the cow decision. The overlap lies in the background region rather than on the animal itself, and the thesis uses this example to argue that CAVLI can identify spurious correlations, shortcut learning, and dataset bias when a classifier depends on irrelevant context (Shukla, 28 Aug 2025).

The CelebA experiment functions as a bias-detection case study. A smile classifier is trained in a biased setting by subsampling training data to induce stronger correlation between female plus smiling and male plus non-smiling. Average CDS on test data is then reported by subgroup: Male Smile 0.004, Female Smile 0.013, Male Non-Smile 0.005, and Female Non-Smile 0.007. The thesis states that the highest average CDS scores were for the female smiling group and the lowest were for the male smiling group, and interprets this as alignment with the existing dataset bias (Shukla, 28 Aug 2025).

4. Relation to other interpretability methods, strengths, and limitations

CAVLI is presented as conceptually intermediate between LIME and TCAV. Relative to LIME, it adds explicit concept semantics; relative to TCAV, it adds localization and per-image analysis; relative to standard saliency methods, it yields concept-aware explanations rather than only pixel-importance maps. The main point of the framework is therefore not a large benchmark comparison, but a structural synthesis of concept representation, local attribution, and counterfactual-style perturbation within a single post hoc procedure (Shukla, 28 Aug 2025).

Several strengths are explicit in the thesis. CAVLI provides local concept-based explanation rather than only class-level concept sensitivity. It combines semantic interpretability from TCAV with spatial interpretability from LIME. It acquires a counterfactual flavor without retraining or generative editing, because superpixel masking is used to simulate alternative local versions of the input. It is also positioned as practical for debugging and auditing, particularly when practitioners wish to determine whether a classifier relies on the “right” concepts or on spurious cues such as background features or protected attributes (Shukla, 28 Aug 2025).

The limitations are equally central. The thesis states that CAVLI assumes that greater spatial overlap between concept-relevant and decision-relevant regions implies stronger concept dependence, and explicitly remarks that this reasoning aligns only with the first rung of Pearl’s ladder of causation, namely association. The framework can therefore report high CDS when two concepts co-locate spatially even if only one is causally relevant. The thesis gives the example that skin color and smiling may overlap spatially, which could produce high CDS for both even if only one drives the decision (Shukla, 28 Aug 2025).

Other limitations follow from the implementation. CAVLI requires predefined concepts and cannot discover or prioritize concepts automatically. It depends on segmentation quality, so poor superpixels may distort both the concept and decision heatmaps. The chapter does not provide many low-level hyperparameters, including the number of superpixels $Y$1, the number of perturbation samples $Y$2, exact LIME kernel parameters, concept-classifier training hyperparameters, or exact segmentation settings. There is no dedicated ablation study, no human evaluation, and no formal localization metric, deletion/insertion metric, or quantitative comparison against attribution baselines such as Grad-CAM or SHAP in Chapter 3 (Shukla, 28 Aug 2025).

The computational profile is described only implicitly. One TCAV concept classifier is required per concept and layer; each image must be segmented into superpixels; many perturbed forward passes are needed for LIME; conceptual sensitivity must be computed for each perturbation; and two local surrogate fits are required per image. The thesis therefore characterizes CAVLI as more expensive than a single saliency map but still practical as a post hoc analysis method, while giving no explicit runtime or efficiency study (Shukla, 28 Aug 2025).

5. Nomenclature and adjacent usages across arXiv-adjacent literatures

Within the supplied arXiv material, the exact name CAVLI is used formally only for the vision-interpretability framework described above. The acronym is not explicitly expanded in the chapter text, but the method title is given as "CAVLI: Using counterfactuals to quantify concept influence on classifier decisions" (Shukla, 28 Aug 2025).

At the same time, the term has nearby but nonidentical associations in other technical literatures. Several ultrasound papers are directly relevant if CAVLI is being used more loosely to mean cavitation localization, cavitation imaging, or cavitation-assisted localized intervention. "Equivalent Time Active Cavitation Imaging" presents ETACI as an active pulse-echo modality that maps a pressure-field surrogate by imaging HIFU-driven microbubble oscillation via equivalent-time sampling, and is explicitly described as highly relevant to the broader topic of cavitation imaging/localization while not being a canonical passive cavitation localization method (Blais et al., 2023). "Equivalent-Time-Active-Cavitation-Imaging Enables Vascular-Resolution Blood-Brain-Barrier-Opening-Therapy Planning" develops BP-ETACI, a vessel-resolved anatomy-linked cavitation mapping method for focused-ultrasound blood-brain barrier opening, and is characterized as directly relevant to CAVLI-style cavitation imaging/localization because it seeks to determine where cavitation-bearing microbubbles are within the vasculature (Desmarais et al., 2023).

A different use appears in intravascular therapy. "A novel forward-looking ultrasound catheter for treating vascular occlusions" is described as particularly relevant to what would now be described as cavitation-assisted localized intravascular intervention—i.e., CAVLI—because it combines catheter-based access, forward-looking ultrasound emission, a guidewire-compatible central lumen, local microbubble delivery, and demonstrated inertial cavitation near an occlusion phantom (Liu et al., 2024). In tumor antivascular ultrasound, "Intravital imaging and cavitation monitoring of antivascular ultrasound in tumor microvasculature" couples real-time passive cavitation detection with intravital optical validation of microvascular injury, and is therefore highly relevant to any CAVLI-like framework understood as cavitation-based monitoring or control of biologically meaningful lesioning (Zhao et al., 2022).

By contrast, some apparent lexical similarities are unrelated. "Vialactea Visual Analytics tool for Star Formation studies of the Galactic Plane" is explicitly not about a separate system called CAVLI; the paper contains no mention of the term, and the most likely intended references in that domain are VLVA or the broader VIALACTEA infrastructure rather than CAVLI (Vitello et al., 2018). A plausible implication is that the acronym currently has an ambiguous editorial footprint across domains, but only one of the supplied works uses it as the formal name of a method.

6. Broader significance and open questions

CAVLI’s broader significance lies in the way it operationalizes concept dependence as a local, inspectable, and numerically summarized quantity. It asks whether, under superpixel masking perturbations, the regions supporting a human-defined concept align with the regions driving the classifier’s output. In doing so it supplies both a scalar diagnostic, the CDS, and a pair of localized heatmaps that show where concept relevance and decision relevance coincide. This combination is especially useful in settings where background context, sensitive attributes, or dataset confounds may influence predictions without being evident from global metrics alone (Shukla, 28 Aug 2025).

The thesis positions CAVLI within a broader counterfactual program for interpretable and fair AI. In that larger structure, CAVLI addresses classifier auditing; later chapters introduce ASAC for adversarial counterfactual fine-tuning of biased classifiers, TIBET for prompt-sensitive bias evaluation in text-to-image models, BiasConnect for causal graphs of intersectional bias, and InterMit for training-free mitigation of intersectional bias. CAVLI is therefore the concept-dependence component of a wider framework that uses counterfactual reasoning to explain, audit, and mitigate bias in both discriminative and generative models (Shukla, 28 Aug 2025).

The open problems identified in the source are technically consequential. The first is the move from association to stronger causal claims. The second is automatic concept discovery or prioritization, since the current framework requires predefined concepts. The third is robustness when concepts spatially overlap or when superpixel segmentation is imperfect. These are not peripheral refinements; they define the boundary between CAVLI as an effective associational auditing tool and a future class of methods that could support stronger causal interpretation of model behavior (Shukla, 28 Aug 2025).

Taken together, the supplied literature supports a narrow and a broad reading of the term. In the narrow and formally established sense, CAVLI is a concept-based local interpretability method for vision classifiers centered on TCAV, LIME, Pearson overlap, and the Concept Dependency Score. In a broader and more inferential sense, the acronym resonates with cavitation imaging and cavitation-assisted localized intervention, but those usages are adjacent rather than terminologically fixed. For arXiv-reading audiences, the most precise definition remains the classifier-interpretability framework introduced in 2025, with neighboring ultrasound literatures serving mainly as a reminder that acronym-level disambiguation is essential.