Explainable & Trustworthy Artificial Intelligence

Beschrijving

Artificial Intelligence (AI) has come a long way since its first use and application many decades ago. The use of AI and Machine Learning have seen an immense uptake in the 21st century. The techniques developed in the domain were and are successfully applied to a wide variety of problems, both in academia, private and public industry. As this domain became more and more established in recent years, new challenges arose.

Artificial Intelligence nowadays are complex and sophisticated algorithms that sometimes make it difficult to understand and interpret the decisions or suggestions of the AI system. Explainable AI puts the following properties on the foreground to deliver trust:

• Gaining trust by explaining for example the characteristics of AI output.
• By explaining an AI technique understanding will increase, allowing to investigate if the technique can be transferred to another domain or problem.
• Informing a user about the workings of an AI model so that there is no misinterpretation.
• Confidence of users can be established by using AI models that are explainable, stable but also robust.
• When explaining AI models issues concerning privacy awareness come into play. Private data should not be exposed by the models.
• It is important that actions can be explained. How have we come to specific outcomes and how could we change them?
• Nowadays, a wide variety of people from different background come into contact AI, it is important that they all understand why the system is behaving in such a manner and offer explanations tailored to their needs.

Programma

1. Introduction

In this first lesson, we give a short recap of the basics, followed by the explanation of some general terms that are used in the domain of explainable and trustworthy Artificial Intelligence. This introduction will end with the definition of the challenges within this domain.

• Recap the basics: AI, ML and statistics
• Different types of ML: white-box & black-box
• Interpretability vs Explainability
• Human Uncertainty vs Model Uncertainty
• Challenges

Teachers: Femke Ongenae & Sofie Van Hoecke

Date: 30 September 2024

2. White box models

While black box models offer higher accuracy, white box models are easier to explain and to interpret, unfortunately this leads to a lesser predictive capacity. In the area of white box models, several different approaches will be highlighted:

• Linear Regression
• Generalized Additive Models (GAMs)
• Decision Trees
• Rule-based systems and fuzzy logic

Teacher: Daniel Peralta Cámara

Date: 7 October 2024

3. Interpretability & Explainability

Machine learning systems build models that learn to automate complex tasks by learning from examples. How to get insights into how these models work depends on the type of algorithm used. Getting insights into how our models work can be done by looking at how the model works in general (interpretability), versus how a specific prediction of the model was computed (explainability). Additional hypothetical “What if” questions can be asked to allow for counterfactual reasoning, adding to the toolkit of explainability methods.

• Interpretability versus explainability
• Counterfactuals
• Model distillation
• Dependency plots
• Saliency maps

Teacher: Yvan Saeys

Date: 14 October 2024

4. Online & Transfer Learning

Training machine learning systems can be done before use, i.e. when training it on a stack of pictures first and asking it to make sense of new pictures later. However, it can also be done during use. In the latter scenario the system gets updated whilst it is being used. Sometimes this is necessary because training data is (partially) becoming available after commissioning of the system. Sometimes a system is pretrained on one dataset and the developer wants to retrain the system in order to solve another but related problem, i.e. using a machine vision system that is trained to detect cats to now detect dogs. The developer thus leverages the effort put into the training of the earlier system, hence requiring less training time for the novel system. These and other relations between datasets, their application in training models and the problems we solve with those will be explained in this lesson.

• Online learning
• Change detection
• Transfer learning & domain adaptation (foundation models)

Teacher: Matthias Feys

Date: 21 October 2024

5. Hybrid AI

The oldest forms of machine learning entail rule engines that were hand programmed. Newer forms entail algorithms searching for connections themselves. The first are great in explaining how they reach their conclusions. The latter sometimes give superior predictions, being a lot less brittle, but lack that explainability. To get the best of both worlds, these approaches are sometimes combined. Moreover, allowing an expert to guide a machine learning system can sometimes lead to yet again superior predictions.

• Data-driven vs expert-based approaches
• Finding synergies in data-driven and expert-based approaches
• Combining expert knowledge and machine learning

Teachers: Femke Ongenae & Sofie Van Hoecke

Date: 4 November 2024

6. Robustness

Machine learning systems are extremely fragile: small modifications to their input data can cause them to produce wildly incorrect outputs. These modifications are usually imperceptible or seemingly harmless, making them hard to detect. Such "adversarial perturbations" undermine the trustworthiness of our systems and may pose safety issues under certain circumstances. This lesson explains these problems and what you can do about them.

• Adversarial learning
• Evasion attacks & defenses
• Learning theory

Teacher: Jonathan Peck

Date: 18 November 2024

7. Uncertainty

The notion of uncertainty is of major importance in machine learning and constitutes a key element of modern machine learning methodology. In recent years, it has gained attention due to the increasing relevance of machine learning for practical applications, many of which are coming with safety requirements. In this regard, new problems and challenges have been identified by machine learning scholars, many of which call for novel methodological developments. Indeed, while uncertainty has a long tradition in statistics, and many useful concepts for representing and quantifying uncertainty have been developed on the basis of probability theory, recent research has gone beyond traditional approaches and also leverages more general formalisms and uncertainty calculi.

• Aleatoric and epistemic uncertainty
• First-order uncertainty representations (probabilistic models, calibration methods, set-based representations, conformal prediction, etc.)
• Second-order uncertainty representations (Bayesian methods, ensemble methods, density-based methods, etc.)

Teacher: Willem Waegeman

Date: 25 November 2024

8. Bias & Fairness

When training machine learning systems, the training data can be biased, leading to unwanted outcomes. For example, an HR system trained on old hospital personnel data might discriminate against women for doctor positions and against men for nurse positions, due to historical gender biases in these roles. This session will explain these issues, how to avoid them, how to measure bias and what the limitations of avoiding it are. Also advanced bias and fairness issues in large language models, and generative AI more generally, will be covered.

• Various notions of fairness & impossibility theorem
• Different types of bias & methods to debias
• Ethical and legal guidelines
• Learning fair models
• Uncovering model bias
• Biases in Large Language Models and generative AI

Teacher: Tijl De Bie

Date: 2 December 2024

9. Privacy

Sometimes the quality of machine learning system outputs and privacy are at odds and need to be balanced. However, there are techniques that allow the training of machine learning systems on privacy sensitive data, without exposing the data itself. Those techniques and relevant regulation on these practices are explained in this session.

• Pseudonymization
• K-anonymity
• Differential privacy
• Regulation

Teacher: Tijl De Bie

Date: 9 December 2024

10. Use cases

During the last session, some specific use cases in the domain of Explainable and Trustworthy AI will be discussed.

Teachers: Tijl De Bie, Matthias Feys, Femke Ongenae & Sofie Van Hoecke

Date: 16 December 2024

More information and subscription

www.ugain.ugent.be/explainableAI