CTC: Contribution to Classification of Complex Features

Abstract

Deep convolutional neural networks have achieved remarkable performance, yet their internal decision-making processes often remain opaque. A key challenge in post-hoc explainability is balancing interpretability and fidelity: overly granular explanations (e.g., at the pixel level) can overwhelm users, while approaches that determine the relevance of aggregated input regions often oversimplify explanations, resulting in a loss of faithfulness to the model's true behaviour. In this paper, we propose a novel framework that (i) modifies the Segment Anything Model (SAM) to identify meaningful complex input features, (ii) introduces a technique, Contribution To Classification (CTC), which employs a modified forward pass to assess the relevance of these features rather than relying solely on pixel-level relevance, and incorporates a scaling mechanism to preserve the contribution signal despite propagating only a subset of activations (iii) demonstrates improved input invariance and sensitivity to meaningful perturbations through extensive evaluations on architectures including VGG, ResNet, Inception, and DenseNet, and (iv) releases \href https://github.com/SophiaKalanovska/Contribution-To-Classification the CTC open-source codebase to facilitate further research.