Skip to content

Explainable trustworthy ai

Explainable & Trustworthy AI#

Overview#

This course focuses on techniques for improving the trustworthiness of machine-learning (ML) based systems, including methods for explaining the decision andensuring the safety of ML models. By the end of this course, students will be able to: - generate explanations from complex models - evaluate the quality ofthese explanations with respect to expectations - formulate properties that aML model should enforce - verify whether these properties hold, and exhibit counter-examples if it is not the case.

Outline#

  1. Introduction to trustworthy AI - 1.5h
    • Reminders on machine learning and deep learning architectures
    • Why do we need trustworthy systems?
    • General overview and definitions
  2. Introduction to formal methods for verification - 1.5h
  3. Explainable AI (XAI) - 9h
    1. Explaining models and evaluating explanations (3h)
      • Taxonomy of explanations
      • Post-hoc explanation methods for models and decisions
      • Designing self-explainable models
      • Properties of explanations and how to measure them
    2. Introduction to formal XAI (1.5h)
    3. Practical work (4.5h)
  4. AI Safety - 6h
    1. Current trends in the evaluation of non-linear ML models (1.5h)
    2. Adversarial machine learning - 1.5h
    3. Practical work (3h)

In brief#

  • Period: semester 9
  • Credits: 3 ECTS
  • Number of hours: 18h
  • Apogée:

Theory: Fundamentals of machine learning (gradient descent) and deep learning architectures (CNN, transformers), notions in computer vision. Practice: development in Python, notions in deep learning frameworks (Pytorch or Tensorflow).

Pedagogical team#