|
|
|||||||||||||||||||
Advanced tools of Artificial Intelligence for explainable and reliable systemsUniversity of Southern DenmarkAdministrative Information
The course will be offered in both online and offline mode. The place of conduction of course in offline mode is The Maersk Mc-Kinney Moller Institute, SDU, Odense.
Name: Smith Kashiram Khare Email: smkh@mmmi.sdu.dk Unit: The Maersk Mc-Kinney Moller Institute, SDU Applied AI and Data Science
Basic understanding of machine learning concepts and Programming skills, preferably in Python. Prior experience with machine learning libraries such as scikit-learn, TensorFlow, or PyTorch is recommended but not required.
The recent tangible introduction of Artificial Intelligence (AI) models to everyday consumer-facing goods, services, and tools raises a number of new technological hurdles and ethical concerns. Among these is the fact that the majority of these models are powered by Deep Neural Networks (DNNs), which, although very expressive and helpful, are often complex and opaque. This special course delves into four key areas of modern AI research: Explainable AI (XAI), Uncertainty Quantification (UQ), Continual Learning, and Model Compression, highlighting how each contributes to model transparency, robustness, adaptability, and efficiency. Through a mix of theory, hands-on exercises, and real-world case studies, learners will gain the skills to design AI systems that balance accuracy, interpretability, reliability, and resource efficiency while addressing ethical and societal considerations. Module 1: Towards Reliable AI and Systemic Implications The module covers the DL workflow from data collection and preprocessing to model training and evaluation. The conceptual balance between accuracy and interpretability, performance versus computational cost, and reliability and trust, alongside ethical, legal, and societal considerations such as transparency and accountability. Module 2: Towards Reliable AI and Systemic Implications This module uncovers understanding the role of explainability. Counterfactual explanations, interpretable models vs. black-box models, tools and frameworks for explainable AI (LIME, SHAP, etc.). The module covers leading XAI tools and libraries, highlighting their practical applications and research advancements. It evaluates model transparency and interpretability using established metrics and experimental studies. Module 3: Continual Learning The challenge of model retraining and data drift, techniques for lifelong and incremental learning, reducing data and compute waste through adaptive learning, and real-world applications of continual learning in dynamic environments. Module 4: Model compression for energy efficient AI Measuring and reducing the carbon footprint of AI models, Model compression, pruning, and quantization for efficiency. Module 5: Uncertainty Quantification in Image Classification This module Covers aleatoric and epistemic uncertainty, methods like Deep Ensembles and Monte Carlo Dropout, relevant metrics such as Expected Calibration Error (ECE), Brier score, and negative log-likelihood, and techniques to visualize uncertainty in image classification predictions. Module 6: Case studies and project work Work on case studies and hands-on experiments to explore applications in sensitive domains like medical imaging or finance. Complete individual or team projects to design trustworthy and efficient AI solutions, highlighting interpretability, reliability, and resource efficiency.
The course will be delivered through an online and offline teaching mode, hands-on lab sessions, and interactive discussions. Participants will have access to project resources, including tutorials, code repositories, and reading materials.
Skills
Competences
Lecture (30%): short theory capsules. Experimental/Lab session (20%): Development of models Teamwork (50%): hands-on lab sessions and practical exercises.
There are as such nor examination conditions, however, students are encouraged to attend the theory sessions or complete class-activities.
Successful completion of the course will require active participation in lab sessions and a final presentation of the project report. Research publications are encouraged but not mandatory Assessment - 7-point grading scale Second examiner – None
PhD-level course intended for students in computer science, data science, AI, mathematics, healthcare informatics, and related disciplines with an interest in AI.
26 hours of interactive sessions 14 hours of hands-on lab sessions and practical exercises 85 hours dedicated to the project report writing and final presentation
Textbooks
Online Courses and Tutorial
Papers
|
|
||||||||||||||||||