همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

AdaLayer-CAM: Adaptive Multi-Layer Method for Visualization and Explanation of CNNs

نویسندگان :

MohammadJavad GhaderiPoor¹ Latifeh PourMohammadBagher² Zahra Rahimi³

1- Allameh Tabataba'i University 2- Allameh Tabataba'i University 3- Allameh Tabataba'i University

کلمات کلیدی :

Convolutional Neural Networks (CNNs)،Explainable AI،Class Activation Mapping (CAM)

چکیده :

Class Activation Mapping (CAM) is one of the standard tools for visualizing the regions of focus in neural networks, showing where the model pays attention during prediction. However, most existing versions such as Grad-CAM, Grad-CAM++, and LayerCAM rely only on a subset of high-level convolutional layers—usually the last layer of the network or the last layer of each stage—and consequently ignore many intermediate and shallow layers that contain valuable spatial and semantic information. In addition, these methods assign equal importance and weight to all selected layers, disregarding the differences in the influence of each layer. In this study, we propose AdaLayer-CAM, a fully adaptive framework that aggregates activation information from all convolutional layers and assigns each layer a weight based on the data and its activation energy. This adaptive weighting makes the layers with stronger discriminative content more prominent while reducing the impact of weaker or noisier layers. Moreover, a soft fusion mechanism replaces the traditional hard-max combination, which gradually and thresholdedly integrates the weighted maps, resulting in more stable and accurate final heatmap localization. The resulting heatmaps preserve the structural precision of shallow layers and the semantic robustness of deeper ones—without requiring any modification to the model architecture or retraining of the network. Experimental results show that AdaLayer-CAM produces class activation maps that are more accurate and less noisy than the original LayerCAM and other commonly used related methods, thereby improving interpretability and localization performance in weakly supervised learning approaches.