| CODE | ARI3205 | ||||||||||||
| TITLE | Interpretable AI for Deep Learning Models | ||||||||||||
| UM LEVEL | 03 - Years 2, 3, 4 in Modular Undergraduate Course | ||||||||||||
| MQF LEVEL | 6 | ||||||||||||
| ECTS CREDITS | 5 | ||||||||||||
| DEPARTMENT | Artificial Intelligence | ||||||||||||
| DESCRIPTION | This study-unit will introduce students to techniques for interpreting deep learning models' behaviour. They will explore what interpretable AI is and why it is essential for modern data science applications employing deep learning models. The study-unit aims to expose students to various types of interpretability techniques, including i) visual attribution methods for interpreting the behaviour of Convolutional Neural Networks, ii) the introduction of Network Dissection Framework, iii) techniques evaluating the semantic similarity of word representations for NLP applications. Moreover, the students will learn to apply a suite of techniques for building interpretable deep learning models. The topics that will be covered include: - Perturbation-based techniques (LIME, SHAP) - Gradients-based techniques (Vanilla, Integrated, Smooth and Guided) - Activations-based techniques (GradCAM and Guided GradCAM) - Techniques for identifying features and concepts (understandable by humans) learned by a Convolutional Neural Network - Techniques for detecting influential instances and adversarial examples - Techniques for visualising and validating the semantic similarity of word embeddings. Study-unit Aims: - Give students the necessary fundamentals of for interpreting deep learning models' behaviour; - Provide the student with the skills to apply interpretable AI techniques for building AI systems targeting real-world problems; - Expose the students to the dominant technologies used for building interpretable deep learning based systems. Learning Outcomes: 1. Knowledge & Understanding By the end of the study-unit the student will be able to: - Outline the various properties of interpretable AI techniques; - Discuss the pros and cons of different approaches for interpreting Convolutional Neural Networks; - Evaluate a repertoire of interpretable AI techniques to be used for interpreting different deep learning models; - Identify influential instances and adversarial exaples and explain their implications on the employed machine learning models. 2. Skills By the end of the study-unit the student will be able to: - Interpret and communicate the outputs of deep learning models; - Determine the appropriate interpretation technique to use in a variety of different contexts; - Build interpretable AI systems based on deep learning for real-world problems. Main Text/s and any supplementary readings: Main texts: - Thampi, Ajay. Interpretable AI: Building explainable machine learning systems. Simon and Schuster, 2022. - Mas铆s, Serg. Interpretable Machine Learning with Python: Build explainable, fair, and robust high-performance models with hands-on, real-world examples. Packt Publishing Ltd, 2023. Supplementary texts: - Linardatos, Pantelis, Vasilis Papastefanopoulos, and Sotiris Kotsiantis. "Explainable ai: A review of machine learning interpretability methods." Entropy 23.1 (2020): 18. - Samek, Wojciech, et al., eds. Explainable AI: interpreting, explaining and visualizing deep learning. Vol. 11700. Springer Nature, 2019. |
||||||||||||
| STUDY-UNIT TYPE | Lecture, Independent Study & Tutorial | ||||||||||||
| METHOD OF ASSESSMENT |
|
||||||||||||
| LECTURER/S | Konstantinos Makantasis |
||||||||||||
|
The University makes every effort to ensure that the published Courses Plans, Programmes of Study and Study-Unit information are complete and up-to-date at the time of publication. The University reserves the right to make changes in case errors are detected after publication.
The availability of optional units may be subject to timetabling constraints. Units not attracting a sufficient number of registrations may be withdrawn without notice. It should be noted that all the information in the description above applies to study-units available during the academic year 2025/6. It may be subject to change in subsequent years. |
|||||||||||||