Review of white box methods for explanations of convolutional neural networks in image classification tasks

Abstract

In recent years, deep learning has become prevalent to solve applications from multiple domains. Convolutional neural networks (CNNs) particularly have demonstrated state-ofthe-art performance for the task of image classification. However, the decisions made by these networks are not transparent and cannot be directly interpreted by a human. Several approaches have been proposed to explain the reasoning behind a prediction made by a network. We propose a topology of grouping these methods based on their assumptions and implementations. We focus primarily on white box methods that leverage the information of the internal architecture of a network to explain its decision. Given the task of image classification and a trained CNN, our work aims to provide a comprehensive and detailed overview of a set of methods that can be used to create explanation maps for a particular image, which assign an importance score to each pixel of the image based on its contribution to the decision of the network. We also propose a further classification of the white box methods based on their implementations to enable better comparisons and help researchers find methods best suited for different scenarios.

Multi Layered Feature Explanation Method for Convolutional Neural Networks

“…Methodology. In [1] and [13] the methodology of evaluation of explanation methods was proposed. It consists in the comparison of the pixel importance maps obtained by the network sensing with the maps expressing human perception of the same visual content.…”

“…The need for explanations of Deep Neural Networks (DNN) decisions has led to an active research in eXplainable Artificial Intelligence (XAI). In the field of pattern recognition for images and videos, explanation of a DNN's decision consists in identifying the set of input pixels that have contributed the most into the decision [1]. A famous example of decisions on a wrong data is given by Ribeiro et al [2]: here, a trained classifier wrongly used the presence of snow as the distinguishing feature between the "Wolf" and "Husky" classes.…”

“…Ayyar et al [1] analyzed the variety of explanations methods highlighting important patterns from the input image. Following their proposed taxonomy, ⋆ Supported by LaBRI.…”

Machine Learning in Cardiovascular Imaging: A Scoping Review of Published Literature

“…Even though DL models are commonly referred to as black boxes, there are ways to visualize the regions of the imaging data to which they contribute the most when making predictions. This is called DL interpretation (or explanation ) mapping [ 49 ]. Despite its uttermost importance, the majority of research employing DL classifier or regressor models did not report interpretation mapping for their models.…”

Weakly supervised Unet: an image classifier which learns to explain itself

“…The copyright holder for this this version posted September 12, 2022. ; https://doi.org/10.1101/2022.09.09.507144 doi: bioRxiv preprint As standard CNN decision functions are not easily invertible (Finzi et al, 2019), voxels' contributions to the image-level predictions are unavailable, leading to the development of several indirect approaches to quantify surrogate voxel-importance measures (Ayyar et al, 2021), of which most may be classified as backpropagation-based and perturbation-based (van der Velden et al, 2022). Class activation mapping approaches (Selvaraju et al, 2020;Simonyan et al, 2014;Zhou et al, 2016) estimate regions' importance by extracting gradients from the final convolutional layers.…”