How to explain the behavior of vision transformers?
This page's goal is to present techniques that can shed light on how Vision Transformers' models (ViTs) operate. We will first have a refresher on the ViTs and how they work. We will develop a simple ViT classifier trained on the 🐶🐱 dataset and use a pre-trained model to efficiently classify the images. The next step is to introduce various methods to visualize the way that the classifier makes specific decisions. These approaches range from visualizing the attention maps to visualizing the query/key and value, but also using the backpropagated gradient similar to gradCAM algorithm. We will make use of PyTorch implementation to demonstrate some of these techniques. At the end of the blog post, there is a simple assignment TODO that you will need to solve as homework.
Gradient-based feature attribution
This page's main focus is to analyze a branch of explainable & interpretable AI (XAI) called posthoc XAI. We will analyze theory, taxonomy, applications, shortcomings of posthoc XAI approaches and apply them on image classification using popular CNN architectures and explain their black box nature. Part of the assessemnet for this tutorial/workshop, will be some research questions that needs be answered by you. These questions can be found all over this blogspot using the TOSUBMIT tag and will be summarized them at the end of the blogspot.
Why should I trust you? Local Interpretable Model-agnostic Explanations
This page's goal is to present a PostHoc feature attribution XAI methodology called LIME (Local Interpretable Model-agnostic Explanations) and demonstrate how it can be used to explain image classification tasks. You will be guided through the code and the results of the LIME algorithm. Part of this workshop will be some research questions that need to be answered by you. These questions can be found in the last part of the tutorials and you will encounter these questions with the TOSUBMIT tag. You will need to submit your answers to these questions in the form of a .pdf file on Canvas. The deadline for submitting your answers is the .. of June 2023.