CLAM: Class-wise Layer-wise Attribute Model for Explaining Neural Networks

Seung-Ho Han, Ho-Jin Choi

Keywords: Interpretable Deep Learning, Explainable AI, Layer-wise Relevance Propagation

TL;DR: Proposing a novel explainable model called CLAM, Class-wise Layer-wise Attribute Model, for explaining deep learning models

Abstract: Deep learning techniques have been actively researched for solving complex and diverse problems, demonstrating high performance across various AI domains. However, the complexity and opaqueness of deep learning-based models often make them "black boxes," leading to concerns about transparency and trustworthiness, which in turn hampers their real-world applicability. To address this issue, numerous studies have been carried out to interpret deep learning models and provide explanations for their predictions or propose interpretable new structures. In this paper, we introduce the Class-wise Layer-wise Attribute Model (CLAM), which aims to provide more accurate and detailed explanations for model predictions in image classification. Specifically, CLAM is designed to work in conjunction with a pre-trained image classification model and an existing interpretable algorithm to learn class-wise layer-wise attributes from the model features. Additionally, when generating a relevance map for new input images, CLAM leverages the learned attribute information to enhance the areas related to the target class thereby improving accuracy. Furthermore, we identify and present the influence of specific samples from the training dataset on the calculated relevance map, offering a higher level of explanation compared to existing methods. To validate the effectiveness of our proposed model, we present quantitative and qualitative experimental results using CUB-200-2011 and ImageNet datasets, along with pre-trained VGG-16 and ResNet-50 image classification models and well-known explainable models.

Primary Area: visualization or interpretation of learned representations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5971