Symposium: 35 Years of Neural Networks and Machine Learning

Symposium: 35 Years of Neural Networks and Machine Learning

On the occasion of Prof. Bert Kappen's retirement, a symposium will be held entitled 35 Years of Neural Networks and Machine Learning.

This summer, I (Prof. Bert Kappen) have officially retired from Radboud University. During the 35 years that I was at the university, the field of machine learning and AI has dramatically evolved from a harmless academic activity to one of the most influential forces in society.

Programme

During this farewell symposium, there will be a number of talks from friends and colleagues that illustrate this transformative period.

• 1.30 pm - Ruedi Stoop (ETH Zurich)

  Ruedi Stoop is professor at the ETH Zurich and University of Zurich as well as Regent’s Prof. at the University of Xi’an in China. Ruedi has been the founder of the Center for Recreational Physics, which promotes good physics in a relaxed and friendly environment. Ruedi has a lifelong interest in complexity science and the miraculous properties of non-linear dynamical systems. He is a central figure in this field and he has organised numerous workshops and conferences, some of the most memorable ones we organized together at his mountain house in the Alps.

  Real-world complexity. Challenges and Solutions.

  Ecology, biology and brain sciences, confront us with real-world data complexity that reaches far beyond what traditional computation approaches can handle, in concepts and computational means. In the past, we either focused on microscopic physics following simplistic behaviour (e.g., Bloch’s laser equations) or on relatively simple macroscopic descriptions lacking a straightforward connection to the microscopic world (e.g., pendulum equations). For systems from both classes, even when starting from measured data, a system description and a corresponding classification is generally achievable (although perhaps not entirely nontrivial).

  At the ‘mesoscopic’ scale in between we deal, however, with an elevated problem level, brought about by inherently complex subsystems that interact in various and variable ways, prominently exemplified by spatiotemporal chaos or biological neural networks. For such cases, the traditional concepts are doomed to fail (typically because of a strongly increased level of underlying dimensionality and non-exponential decay of correlations). We propose and demonstrate a novel approach that solves these issues. Finally, we discuss how failure of recognising these issues may have already led to ill-designed research directions.

• 1.55 pm - Alex Khajetoorians (RU)

  Alex Khajetoorians is professor of condensed matter physics at the Radboud University. From the beginnings of the Bits and Brains initiative, Alex and I have shared a common interest in how to build computer architectures that use neural networks and learning, similar to the human brain. This has led to a proof of concept to use individual atoms as neural computing elements.

  What can we ‘learn’ from atoms?

  What can we ‘learn’ from atoms? There is a quest to create materials that are computers, where there’s no longer a distinction between hardware and software. Often, these concepts are routed in mimicking the basic computational properties of the brain and trying to realise them in the physical behaviour of materials. Efforts toward this end requires understanding how concepts like memory and association, can be linked to fundamental physical concepts.

  Here, I will review the ideas linked toward this new paradigm in computing, routed in fundamentals studies based on the idea of “let the physics do the work for you.” I will then demonstrate a bottom-up platform we have developed, based on the concept of atomic orbital memory, which allows us to explore the fundamental physics related to this paradigm. In this discussion, I will review the link to multi-well energy landscapes and the dynamics of glasses. I will then show how the stochastic dynamics of coupled atoms directly mimic the perennial Boltzmann machine. I will conclude with an outlook on these concepts and future avenues in brain-inspired computing and their connections to quantum technologies.

• 2.20 pm - Wim Wiegerinck (Smart Research Nijmegen)

  Wim Wiegerinck is my long term colleague at Smart Research. This company was founded by Stan Gielen, Tom Heskes and me in 1997 and Tom was the first director. The first product was a neural network for prediction of newspaper sales that was sold to the major Dutch newspaper de Telegraaf. Over the years, Wim and I have worked on many projects together, such as the Japanese Real World Computing program from 1992-2002 and the medical diagnostic decision support system Promedas.

  SNN and missing person identification through DNA

  Wim will review some of the history of neural networks and machine learning in Nijmegen, starting with SNN in 1989. Since 2009, Smart Research started to develop Bonaparte, a software program for identification of missing persons based on DNA. Bonaparte has been used by the Dutch Forensic Institute (NFI) for the identification of the victims of the MH17 plane crash over Ukraine and has been essential in solving various cold criminal cases, such as the Marianne Vaatstra case. Other customers of Bonaparte are Interpol for their international missing person program I-Familia, the Australian police force for their criminal investigations and the Spanish government for the identification of victims form the Spanish Civil War.

• 2.45 pm - Taylan Cemgil (Google Deepmind London)

  Taylan Cemgil was one of my PhD students, he graduated in 2009. Since 2019 he joined Google Deepmind in London, working on evaluation of AI systems. He is interested in statistical machine learning methodologies, factuality, human-AI collaboration and scientific applications of large language models. Before joining GDM, he was a professor of Computer engineering in Bogazici University, Turkey, leading a research group working on probabilistic machine learning, tensor decompositions and Monte Carlo methods and collaborating with industry on several applications such as audio and music processing, anomaly detection, tracking or customer analytics. Taylan has the habit to write computer programs for the purpose of improving his understanding of whatever new problem in machine learning he encounters. These programs are so good, that we are still using some of them today!

  Model Evaluations under Uncertain Ground Truth

  AI systems undergo thorough evaluations before deployment, validating their predictions against a ground truth which is often assumed to be fixed and certain. However, in many domains, such as medical applications, the ground truth is often curated in the form of differential diagnoses provided by multiple experts. While a single differential diagnosis reflects the uncertainty in one expert assessment, multiple experts introduce another layer of uncertainty through potential disagreement.

  In this talk, Taylan will argue that ignoring uncertainty leads to overly optimistic estimates of model performance, therefore underestimating risk associated with particular diagnostic decisions, leading to unanticipated failure modes. We propose a statistical aggregation approach, where we infer a distribution on probabilities of underlying medical condition candidates themselves, based on observed annotations. This formulation naturally accounts for the potential disagreements between different experts, as well as uncertainty stemming from individual differential diagnoses, capturing the entire ground truth uncertainty.

  We conclude that, while assuming a crisp ground truth can be acceptable for many AI applications, a more nuanced evaluation protocol should be utilized in medical diagnosis. If time permits, I will also cover some work, based on conformal methods that can provide statistical guarantees. (Based on joint work with David Stutz, Melih Barsbey, Alan Karthikesalingam, Arnaud Doucet and many others.)

• 3.40 pm - David Barber (University College London)

  David Barber was for two years a postdoc in my lab. He is currently Director of the Centre for Artificial Intelligence at University College London. He studied Mathematics and Cambridge and has a PhD in Statistical Mechanics from Edinburgh University. His main interest is in using probability as a tool to model AI. He cofounded the startups re:infer (acquired by UiPath) and Humanloop (acquired by Anthropic).

  From Physics to Philosophy : The Quest for AI

  AI has a long history and different communities have played their part in the evolution of the field. David will describe some early inspirations from physics and how complexity arising from simple rules has been a common theme. He will also discuss the current models behind Generative AI and whether the standard reductionist paradigm is appropriate for understanding why they work. Finally, where is this all going?

• 4.05 pm - Paavo Pylkkanen, Marilu Chiofalo and John Shawe Taylor

  The joy of slow science

  The COVID-19 period gave me the opportunity to reflect on my scientific career. I realized that my secret hoped that by exploring the computational and physical principles of the brain, one could gain some understanding of what it means to be a sentient human being had failed miserably. I started reading authors that had thought about these issues, and others that claim to have a theory of consciousness. Through these explorations, I met Paavo Pylkkanen (Helsinki University) and Marilu Chiofalo (University of Pisa), and COVID-19 gave us the opportunity for long weekly Zoom meetings to discuss scientific theories of consciousness and many other topics. Some time later, John Shawe Taylor (University College London) joined our discussions. Paavo, Marilu and John will reflect on this special period and what it teaches us about doing science.

• 4.20 pm - Bert Kappen

  The limits of science

5.00 pm - Reception

When Friday 31 October 2025, 1:30 pm - 6 pm Location Red Room (01.201) Trigon Organisation Donders Centre for Neuroscience Registration deadline Friday 24 October 2025