Keynote speaker

Abstract: The development of modern intelligent systems and the success of their applications are linked with using of increasingly complex machine learning models that are considered as “black boxes” — they neither provide human global information about their internal model logic nor local reasons for the decision made for the given instance. This has motivated the growing research interest in explainable artificial intelligence (XAI). Unfortunately, in XAI there is now a trend to constantly introduce many new methods, while the assessment of their usefulness is too limited (mainly by placing too much emphasis on the fidelity of the explanation with respect to the original model). Moreover, the number of comparative studies of XAI methods is too limited and the spectrum of studied measures is too narrow. In our opinion more research attention should be devoted to experimental evaluations, both quantitative (measures for more automatic studies) and qualitative (human or application grounded) ones and to the appropriate interaction with human experts. This talk will firstly review the current proposals for evaluating local explanations and the guidelines for obtaining human appropriate feedback.
Next, using the illustrative example of the DetoxAI system – an image recognition deep neural network focused on ML fairness issues, we will show how XAI methods can help detect undesirable biases that erroneously affects the system’s performance, identify the part of the neural network responsible for it, and “debiase” by concept unlearning. As final fairness measures are improved, this demonstrates usefulness of applying XAI methods to this specific real-world task. 

However, we should be aware that unlike standard ML learning task, the XAI field still lacks datasets with well-defined ground-truth explanations or at least sufficient annotations, so researchers must manually review the generated evaluation. On one side, we can postulate development such benchmark datasets. On the other hand—a still challenging task is - how to involve humans in evaluating the usefulness of proposed solutions and to link this to their trust in performance and predictions of ML systems. This is still an open issue, as psychologically well based human evaluation still remains rare. We will illustrate the challenges in human studies with a project comparing the effectiveness of various concept-based methods for explaining neural network predictions for images.

Yet another topic discussed in this talk includes using XAI for explaining changes in data streams, where data distributions and models evolve over time – which is commonly studied as concept drift. Recall, that most of standard XAI methods are designed for static settings, where models are trained once and explanations are generated for fixed data and model. Applying such static XAI to evolving data leads to failures, as explanations may become stale, inconsistent over time and misleading. They are treated as static artifacts, while data and model evolve. This talk will present the speaker's latest original works and experiences with adapting prototypes and counterfactuals for this context, and to better characterize causes of the concept drift. Finally, it is related to challenges of evaluation of such aspects. It requires new metrics and evaluation scenarios which takes into account temporal dynamical of data, model and explanation, which are considered all together.

Bio: Jerzy Stefanowski works as a full professor at Poznan University of Technology, Institute of Computing Science. He received his Ph.D, and Habilitation degrees from the same University. In 2021 he was elected as a member of Polish Academy of Science, where he also plays a role of a Chair of Scientific Council of Institute of Computer Science (IPI PAN) in Warsaw and vice-president of Thematic Committee of PAN on Computer Science. Since 2022 he is leading the Machine Learning Lab at Poznan University of Technology.

His research interests include data mining, machine learning and intelligent decision support. Major results are concerned with: ensemble classifiers, learning from class-imbalanced data, online learning from evolving data streams, explainable AI, induction of various types of rules, data preprocessing, generalizations of rough set theory, descriptive clustering of texts and medical applications of data mining. He is the author and co-author of over 170 research papers and 2 books, which are highly cited. He was a visiting professor or researcher in several universities, mainly in France, Italy, Belgium, Spain and Germany

In addition to his research activities he served in a number of organizational roles such as vice-president of Polish Artificial Intelligence Society (2014-2026); co-founder and co-leader of Polish Special Interest Group on Machine Learning. Moreover, he is the Editor in Chief of Foundations of Computing and Decision Science journal since 2012.

More details could be found at (http://www.cs.put.poznan.pl/jstefanowski/)