An Interpretability Optimization Method for Deep Learning Networks Based on Grad-CAM

Abstract

In the modern era, the optimization of energy utilization in both industrial production and everyday life represents a critical challenge. This challenge underscores the imperative for innovative technological solutions, with artificial intelligence (AI) standing at the forefront of this transformative wave. Among the myriad advancements, high-performance deep learning frameworks have notably advanced, propelling the widespread adoption of computer vision technologies. Despite these advancements, the quest for enhanced interpretability of deep learning networks remains a concern, as it holds the key to unlocking more transparent AI systems. To address this issue, our study proposed an innovative method called information activation mapping (IAM), aimed at enhancing the interpretability of deep learning classification networks through the advanced gradient-weighted class activation mapping-based approach. By refining and extending existing methods, our technique produces highly detailed highlight maps that illuminate critical areas influencing decision-making processes within the network. This breakthrough not only elevates the network’s interpretability but also serves as a cornerstone for development of more efficient and targeted datasets. Our comprehensive analysis encompassed both the qualitative and quantitative evaluation of data quality, alongside a thorough efficiency analysis of the networks and datasets involved. By doing so, we demonstrate that our proposed IAM method significantly improves the interpretability of deep learning networks, thereby promoting more accurate and effective data processing strategies. This research has profound implications for future of AI, particularly in the realms of sustainable and low-carbon AI of Things (AIoT) systems, by promoting more energy-efficient data processing and contributing to the advancement of eco-friendly technological solutions.