Julius von Kügelgen is an (incoming) postdoc with Jonas Peters at ETH Zürich doing research at the intersection of causal inference and machine learning. He pursued his PhD under the supervision of Bernhard Schölkopf at the Max Planck Institute for Intelligent Systems in Tübingen and Adrian Weller at the University of Cambridge. During his PhD, he interned at Amazon and visited Columbia University and UC Berkeley. Previously, he studied Mathematics at Imperial College London and Artificial Intelligence at UPC Barcelona and at TU Delft.

Causal models provide rich descriptions of complex systems as sets of mechanisms by which each variable is influenced by its direct causes. As such, they support reasoning about manipulating parts of the system and capture a whole range of interventional distributions. Causality therefore holds promise for addressing some of the open challenges of artificial intelligence (AI), such as planning, transferring knowledge in changing environments, or robustness to distribution shifts. In this lecture, I will provide an introduction to the field of causal inference from a machine learning perspective. First, we will cover different types of causal models and how to use them to compute quantities of interest ("causal reasoning"). We will then discuss key assumptions and how to leverage them to learn causal models from data ("causal discovery"), including in temporal settings. Lastly, we will turn to connections between causality and machine learning, including the recently emerged field of causal representation learning, which hopes to overcome some of the limitations of traditional methods.

Gido van de Ven is a MSCA postdoctoral fellow at the KU Leuven (Belgium), where he performs research at the intersection of deep learning, computational neurosience and cogntive science. In his research, Gido van de Ven uses insights and intuitions from neuroscience to make the behavior of deep neural networks more human-like. In particular, he is interested in the problem of continual learning, with generative models a principal tool to address this problem. Previously, for the doctoral research, Gido van de Ven used optogenetics and electrophysiological recordings in mice to study the role of replay in memory consolidation in the brain.

Incrementally learning new information from a non-stationary stream of data, referred to as 'continual learning', is a key feature of natural intelligence, but an open challenge for deep learning. For example, standard deep neural networks tend to catastrophically forget previous tasks or data distributions when trained on a new one. Enabling these networks to incrementally learn, and retain, information from different contexts has become a topic of intense research. In the first part of this lecture, I introduce the continual learning problem. After covering key terminology, I review the popular benchmarks and setups currently used in the literature. In particular, I discuss three different types of continual learning, that each have their own challenges: task-, domain- and class-incremental learning. I also cover the distinction between task-based and task-free continual learning. I end this part of the lecture with practical examples to illustrate why continual learning is so important. In the second part of the lecture, I review approaches that have been proposed to address the continual learning problem. I do this at the level of computational strategies, covering the following: context-specific components, parameter regularization, functional regularization, replay, template-based classification, and optimization-based approaches. For each strategy I highlight some representative example methods, and code is available to gain hands-on experience.

Anna Hedström is a third-year PhD candidate at Technische Universität Berlin where she is a part of the independent research group, Understandable Machine Intelligence Lab (UMI Lab), focusing on eXplainable AI (XAI) and Interpretable Machine Learning (IML) topics. Her current research interests include deep learning, software for IML and in particular, evaluation-centric XAI. She has been published and/ or presented her work at AAAI, JMLR, TMLR, ICLR and XAIA NeurIPS and reviewed for prestigious venues such as IEEE TNNLS and ECAI. AH received her BSc from University College London (UCL) and her MSc degree in ML from the Royal Institute of Technology (KTH). She has previously held different research- and ML positions at various companies and AI start-ups and is the main developer of the popular open-source library Quantus.

In this keynote series, we explore the rapidly evolving field of Explainable AI (XAI). First, beginning with a foundational overview, we trace the evolution of XAI methods from early neural network models to the recent transformer-based architectures, assessing their strengths and failure modes. Second, central to our discussions will be the critical theme of XAI evaluation, a previously understudied area that has caused confusion about which explanation methods work and under what conditions. We will concentrate on the fundamental challenge of XAI evaluation: how to establish and verify ground truth and review the community-driven approaches to it. Third, we will introduce Quantus in a hands-on practical session. Quantus is a toolkit for evaluating neural network explanations, exemplifying the practical application of XAI theory. Our goal is to equip the attendees with a deep understanding of both theoretical and practical aspects of XAI, providing a balanced view of the challenges and opportunities that characterise the current state of the field.

Dr. Naser Damer is a senior researcher at the competence center Smart Living & Biometric Technologies, Fraunhofer IGD. He received his master of science degree in electrical engineering from the Technische Universität Kaiserslautern (2010) and his PhD in computer science from the Technischen Universität Darmstadt (2018). He is a researcher at Fraunhofer IGD since 2011 performing applied research, scientific consulting, and system evaluation. His main research interests lie in the fields of biometrics, machine learning and information fusion. He published more than 50 scientific papers in these fields. Dr. Damer is a Principal Investigator at the National Research Center for Applied Cybersecurity CRISP in Darmstadt, Germany. He serves as a reviewer for a number of journals and conferences and as an associate editor for the Visual Computer journal. He represents the German Institute for Standardization (DIN) in ISO/IEC SC37 biometrics standardization committee.

Dr. Fadi Boutros is a scientific researcher at the Fraunhofer IGD and a principal investigator at the National Research Center for Applied Cybersecurity ATHENE, Germany. Fadi received his Ph.D. in computer science from TU Darmstadt (2022) and a master's degree in "Distributed Software Systems" from TU Darmstadt (2019). Also, he is participating in the Software Campus program, a management program of the German Federal Ministry of Education and Research (BMBF). He authored and co-authored several conference and journal papers. His main research interests lie in the fields of biometrics, machine learning, and efficient deep learning. For his scientific work, he received several awards, including the CAST-Förderpreis 2019 award, the IJCB 2022 Qualcomm Audience Choice Award, and the 2022 EAB Biometrics Industry Award from the European Association for Biometrics (EAB) for his Ph.D. dissertation.

The availability of sufficient and diverse data to train neural networks is essential for such networks to operate properly in real applications. Data collection, use, re-use, and share is not only expensive in term of effort and actual cost, but can be impossible in some cases. Such processes faces more challenges when restricted by legal and ethical frameworks, such as the case when dealing with personal data. The raise in generative models presents the chance to explore generating data, with more granular diversity than the data used to train it. Thus, can help generating data more suitable to train neural networks both in terms of availability and representation of reality. This session will focus on motivating different use cases of synthetic data in training neural networks, with a focus on generation mechanisms specifically designed to produce data with training data properties. Given the strong legal and ethical motivations, as well as the unique requirements, most examples will focus on generating synthetic faces and training face recognition models. The audience is expected to gain insights into generating and using synthetic data for the purpose of training neural networks in this relatively early stage of technology adaption.

Jason Eshraghian is an Assistant Professor and Fulbright Fellow at the Department of Electrical and Computer Engineering at the University of California, Santa Cruz. He works on neuromorphic chip and algorithm design, primarily trying to understand how to bridge the gap between artificial intelligence and natural intelligence. He is the developer of snnTorch, a Python library used to train and model brain-inspired spiking neural networks which has been downloaded over 100,000 times. He has received several IEEE Best Paper and Best Live Demo Awards for his work in neuromorphic computing.

The brain is the perfect place to look for inspiration to develop more efficient neural networks. Our brains are constantly adapting, our neurons processing all that we know, mistakes we've made, failed predictions - all working to anticipate what will happen next with incredible speed. Our brains are also amazingly efficient. Training large-scale neural networks can cost more than millions of dollars in energy expense, yet the human brain does remarkably well on a power budget of 20 watts.

One of the main differences with modern deep learning is that the brain encodes and processes information as temporal spikes rather than continuous, high-precision activations. This presentation will dive into the intersection between neuroscience and deep learning. We will explore how spiking neurons and learning rules derived from neuroscience intersect with deep learning, and how the Python package Eshraghian developed - snnTorch - spans across the neuroscience, machine learning, and hardware abstractions to drive forward the next generation of deep learning algorithms.