Module description
- Explainable AI
| Number |
eai
|
| ECTS | 3.0 |
| Specification | Understand, apply, and evaluate models using XAI methods. |
| Level | Advanced |
| Content | Explainable AI (XAI) is a collection of methods dedicated to comprehending and ex-tracting insights from deep learning models. This insight can serve multiple purposes, from ensuring model robustness and eliminating biases to advancing scientific knowledge. As machine learning becomes increasingly integrated into society, the demand for model transparency and accountability from both private companies and governmental organizations grows, highlighting the importance of skilled XAI practitioners. During this module, students will acquire the competencies necessary to extract human-level explanations of the decision-making processes underscoring deep learning models across several data modalities. They will be able to assess the model's robustness and communicate their findings effectively to business stake-holders, regulators, and end-users. |
| Learning outcomes | LO1: XAI Primer (Tabular) Students will be familiar with the vocabulary and fundamental concepts of XAI. They will comprehend the distinction between local and global explanations, intrinsic and post-hoc explainability, as well as model-specific and model-agnostic methods. Students will obtain a high-level overview of the major methodological categories in XAI, including: attributions, counterfactuals, and surrogate models. They will analyze, evaluate, and interpret biases in tabular data using several plot-based visualization methods such as PDPs and ICEs. The student will understand how correlations between features complicate the task of XAI, and how Shapley Values offer a theoretical backbone and solution for XAI. L02: Local model Agnostic (Text) Students will utilize two powerful model- and data-agnostic techniques: Local Interpretable Model-Agnostic Explanations (LIME) and Counterfactuals, enabling them to analyze and interpret text-based sentiment data. They will grasp the theory behind both techniques and comprehend their advantages and disadvantages, as well as their domain of applicability. Moreover, they will explore the effects that hyperparameters and interpretable representations can have on the quality of explanations. Additionally, students will know how to apply and interpret a transformer models attention patterns using the open-source tool “BertViz”. L03: Path Integration techniques (Image) Students will become familiar with a set of path integration methods such as Integrated Gradients, XRAI, and Expected Gradients which aim to apply the Shapley value concept to images, as well as model specific techniques such as CAM and Grad-CAM. They will grasp the theory, strengths, weaknesses, and applicable domains of each technique, and be able to identifying and evaluating discriminant region of images. |
| Evaluation | Mark |
| Built on the following competences | Deep Learning |
| Modultype | Portfolio module |
Diese Seite teilen: