Year: 2025
Type: dissertation
Abstract: Artificial intelligence and machine learning (ML) are characterized by rapidly advancing research. ML revolves around developing systems that learn from data and perform tasks without specific instruction. Today, neural networks (NNs), especially convolutional neural networks (CNNs), prevail as leading ML methods proven effective at computationally complex tasks, including image classification, object detection, and language processing. A significant challenge in wider application of NNs is the inherent lack of transparency in how their predictions are processed resulting from their complicated functionality and structure. Alleviating the transparency problem of NNs constitutes the purpose of explainable ML which involves developing methods that explain the logic of a trained ML model. This study investigates visual explanation methods for CNNs that produce heatmaps depicting the importance of data features to the model’s prediction. This literature review inspects the mechanics of CNNs, key concepts of explainable ML, and the value of visual explanations. The study then compares two state-of-the-art explanation methods: gradient-weighted class activation mapping (Grad-CAM) and layer-wise relevance propagation (LRP). They feature prominently in research literature and prevail as clear and applicable solutions to explain CNN predictions. The objective of this study is to formulate a literature-based paradigm to evaluate the quality of Grad-CAM and LRP explanations. Research demonstrates that Grad-CAM excels at class-specificity while remaining computationally lightweight and applicable to most CNN architectures. Its explanations, however, lack fine details, and are not produced according to the model’s predictive logic. Conversely, LRP produces detailed pixel-level distributions and exhibits greater fidelity to the ML model. Compared to Grad-CAM, LRP’s main disadvantages include lower localization accuracy, and increased computational complexity. Consequently, Grad-CAM is more suited to applications featuring large datasets and several distinct classes, whereas LRP suits applications where sensitivity to minute details is required.
Source: Aaltodoc (Aalto University)
Author: Kumpulainen, Joonatan
Institution: Aalto University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7016079206
