Visual-TCAV: Concept-based Attribution and Saliency Maps for Post-hoc Explainability in Image Classification

Researchers introduce Visual-TCAV, a novel explainability framework for image classification that combines concept-based and saliency-based methods to provide both local and global interpretations of CNN predictions. The method demonstrates improved faithfulness compared to existing approaches like TCAV, bridging a gap between understanding where networks recognize concepts and how those concepts contribute to specific predictions.

Visual-TCAV addresses a fundamental challenge in deep learning interpretability: explaining how convolutional neural networks arrive at their decisions. While CNNs excel at image classification tasks, their internal decision-making processes remain largely opaque. Existing explainability methods fall into two camps—saliency methods that highlight relevant image regions but lack conceptual grounding, and concept-based approaches like TCAV that measure sensitivity to human-defined concepts without spatial localization or instance-level attribution.

This research builds on decades of work in neural network interpretability, from early gradient-based visualization techniques to more recent concept-based frameworks. The integration of Concept Activation Vectors with Integrated Gradients represents a methodological advance that enables practitioners to answer both "where" and "how much" questions simultaneously. The controlled experiments demonstrating better ground truth alignment than TCAV suggest the method produces more reliable explanations, which carries implications for high-stakes domains.

For practitioners deploying image classification systems in regulated industries—healthcare, autonomous vehicles, biometric systems—improved explainability tools directly reduce compliance burden and increase model transparency. The public code release accelerates adoption across research and industry. Developers can now construct more trustworthy systems while auditors gain better tools for validation.

The framework's implications extend beyond academic interest. As regulatory pressure mounts for AI interpretability, methods that reliably explain model behavior become competitive advantages. Future developments may focus on scaling Visual-TCAV to larger models, extending it to other domains beyond images, and integrating it into production ML pipelines for systematic model auditing.

• →Visual-TCAV merges local saliency maps with global concept-based explanations, addressing limitations of existing single-approach methods.
• →Controlled experiments show the method achieves better ground truth alignment than TCAV for explaining CNN predictions.
• →The framework enables practitioners to determine both where concepts appear in images and their quantitative contribution to predictions.
• →Open-source code availability accelerates adoption in research and industry applications.
• →Enhanced interpretability tools support regulatory compliance and trust-building in high-stakes image classification deployments.