Explainable & Trustworthy AI
AI-algoritmen zijn complex en zeer geavanceerd. Het is voor de eindgebruiker soms moeilijk om de beslissingen of de suggesties die door het systeem gemaakt worden, te begrijpen en te interpreteren. Deze opleiding bekijkt de verschillende aspecten die te maken hebben met vertrouwen.
Praktische info:
Leertraject
Inschrijven?
- Inschrijvingen: tot 20 mrt. 2023
- Voorwaarden: Voortgezette opleiding in computerwetenschappen of een equivalente programmeerervaring met Python of een gelijkaardige programmeertaal
- Prijs: €1.650 on-campus / €1.400 online
georganiseerd door:
AI-algoritmen zijn complex en zeer geavanceerd. Het is voor de eindgebruiker soms moeilijk om de beslissingen of de suggesties die door het systeem gemaakt werden, te begrijpen en te interpreteren.
Artificiële Intelligentie (AI) heeft een lange weg afgelegd sinds enkele decennia geleden voor het eerst werd toegepast. AI en machine learning zijn ondertussen heel populair en de technieken worden met succes toegepast op een breed scala aan problemen, zowel in de academische wereld, de industrie als de publieke sector. Naarmate dit domein de laatste jaren meer en meer ingeburgerd raakte, ontstonden er ook nieuwe uitdagingen omdat de werking en uitkomsten niet altijd door mensen gevolgd kunnen worden.
De opleiding ‘Explainable AI’ bekijkt de verschillende aspecten van AI die te maken hebben met vertrouwen:
- Vertrouwen winnen door bijvoorbeeld de kenmerken van AI-output uit te leggen.
- Door een AI-techniek uit te leggen zal het begrip toenemen, waardoor kan worden onderzocht of de techniek kan worden overgedragen naar een ander domein of probleem.
- Een gebruiker informeren over de werking van een AI-model zodat het niet verkeerd begrepen wordt.
- Het vertrouwen van gebruikers kan worden vastgesteld door gebruik te maken van AI-modellen die verklaarbaar, stabiel maar ook robuust zijn.
- Bij het uitleggen van AI-modellen spelen vraagstukken rond privacy een rol. Privégegevens mogen niet door de modellen worden onthuld.
- Het is belangrijk dat handelingen en acties kunnen worden uitgelegd. Hoe zijn we tot specifieke resultaten gekomen en hoe kunnen we deze veranderen?
- mensen met verschillende achtergronden komen met AI in aanraking, het is belangrijk dat ze allemaal begrijpen waarom het systeem zich zo gedraagt en dat een uitleg gegeven wordt, die is afgestemd op hun behoeften.
Elke module binnen deze cursus richt zich op een ander onderwerp binnen het domein van verklaarbare en betrouwbare AI, variërend van white-box-modellen over hybride technieken tot aspecten van privacy. Deze cursus wordt georganiseerd in 10 verschillende modules.
Programma
Introduction, 20 March 2023
In this first module, 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
Teacher: prof. dr. ir. Sofie Van Hoecke, Department of Electronics and Information Systems, Ghent University & prof. dr. Femke Ongenae, Department of Information Technology, Ghent University
Global vs Local Interpretability, 27 March 2023
Machine learning systems help us connect cause and effect in complex data sets. How we can make those interpretations depends on the type of algorithm used. The level of interpretability we can reach will also differ depending on whether we look at how the model predicts in general, versus how a specific prediction of the model was computed. When interpretability is good enough, we can use hypothetical questions like “What if instead this would be the case?”, adding to the usefulness of your system. This module explains how to increase and measure those interpretabilities.
- Global Interpretability
- Local Interpretability
- Counterfactuals
- Model distillation
- Dependency plots
- Evaluation methods
Teacher: Arne Gevaert, Department of Applied Mathematics, Computer Science & Statistics, Ghent University
White box models, 17 April 2023
In this module, focus is given to 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
- Decision Trees and Rule Sets
- Generalized Additive Models (GAMs)
Teacher: dr. Daniel Peralta Cámara, Department of Information Technology, Ghent University
Saliency mapping, 24 April 2023
To help explain how a neural network reached a certain conclusion, certain visualizations of its reasoning can be useful. One type of visualization is a heatmap showing which areas of a photo contribute most to how a system has labeled it. These visualizations are explained in this session.
- Convolutional Neural Networks (CNNs) & interpretation
- Backpropagation
- Other gradient-based methods
Teacher: dr. Arne Gevaert, Department of Applied Mathematics, Computer Science & Statistics, Ghent University
Hybrid AI, 8 May 2023
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. This module explains how.
- Data-driven vs expert-based approaches
- Finding synergies in data-driven and expert-based approaches
- Combining expert knowledge and machine learning
Teacher: prof. dr. ir. Sofie Van Hoecke & prof. dr. Femke Ongenae, Department of Information Technology, Ghent University
Robustness, 15 May 2023
The output of a machine learning system depends on the data used as input. Often the needed amount and structure of that data is overlooked. However, machine learning systems can be combined to generate additional data or to finetune each other. Nevertheless, malicious additions to your training data can corrupt your system and even a well-trained system can be deceived. This module explains these issues and what you can do about them.
- Data quality & quantity
- Adversarial learning
- Poisoning attacks & defenses
- Learning theory
Teacher: Jonathan Peck, Department of Applied Mathematics, Computer Science & Statistics, Ghent University
Online & Transfer Learning, 22 May 2023
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 module.
- Online learning
- Change detection
- Transfer learning & domain adaptation (foundation models)
Teacher: Jan Van Looy & Matthias Feys, ML6
Bias & Fairness, 5 June 2023
When training machine learning systems, the training data can be biased, leading to unwanted outcomes, i.e. an HR system trained on old hospital personnel data now stating that women might be unlikely good candidates for doctor positions. This module will explain these issues, how to avoid them, how to measure bias and what the limitations of avoiding it are.
- Various notions of fairness & impossibility theorem
- Different types of bias & methods to debias
- Ethical guidelines
- Learning fair models
- Uncovering model bias
Teacher: Prof. Tijl De Bie, Department of Electronics and Information Systems, Ghent University
Privacy, 12 June 2023
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 module.
- Pseudonimization
- K-anonymity
- Differential privacy
- Regulation
Teacher: Prof. Tijl De Bie, Department of Electronics and Information Systems, Ghent University
Use cases, 19 June 2023
During this module, some specific use cases in the domain of Explainable and Trustworthy AI will be discussed.