Visualizing interpretations of deep neural networks

• Main Author: Ta, Quynh Nga
• Other Authors: Li Boyang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/166663
• Institution: Nanyang Technological University

The evolution of Convolutional Neural Networks and new approaches like Vision Transformers has led to better performance in computer vision. However, deep neural networks lack transparency and interpretability, leading to consequences in critical applications. Visualizing deep neural network interpretations can provide insights into decision-making, identify biases and errors, and reveal potential limitations in the model or training data. This area of research is significant for enhancing the transparency, interpretability, and trustworthiness of deep neural networks and facilitating their application in critical domains. This project aims to create a web application to facilitate the interpretation of the ConvNeXt model, a state-of-the-art convolutional neural network. The application implements three techniques: Maximally activating image patches, Feature attribution visualisation with SmoothGrad, and Adversarial perturbation visualisation with SmoothGrad. Maximally activating image patches help users understand what patterns maximally activate a channel in a layer. Feature attribution visualisation with SmoothGrad highlights the pixels that are most influential for the model's prediction. Adversarial perturbation visualisation with SmoothGrad allows users to explore how the model reacts when the input image is perturbed. The results of experimentation of interpretability techniques on the ConvNeXt model will also be discussed in this report.