Year: 2025
Type: article
Abstract: Explainable Artificial Intelligence (XAI) is essential to investigate the biased process of decision-making of deep learning models, particularly when the decision concerns a high-stakes vision outcome. Although Vision Transformers (ViTs) have reached the state-of-the-art performance levels, they are using complex self-attention strategies that hide the rationale. The existing explanation techniques mostly localize the influential pixels, but rarely do they identify a higher-level factor that can be understood by humans or give a causal relationship. This paper presents Concept-based Sobel Analysis (CSA), a model that is intended to produce conceptual-level, causally-based explanations of ViTs. The methodology uses Non-negative Matrix Factorization (NMF) of intermediate feature representations to discover semantically consistent latent concepts and uses Sobol variance decomposition to measure the causal impact of each latent concept on model predictions, both main effects and higher-order interactions. The experiments on a Pyramid Vision Transformer (PVT) show that the explanations generated by CSA are more stable, interpretable, and causally faithful than those of correlation-based baselines. CSA provides a rigorous way to debug and bias detect, and develop trustworthy AI systems by ranking concepts by their contribution to the variation in the output.
Institution: Chongqing University of Posts and Telecommunications
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7127050500