The recent "generative turn" in artificial intelligence shows that an increasing range of intelligent behaviours can emerge in machines. At the [HumanAI]volution conference in April 2026, we will bring together European researchers from computer science, biology, behavioural sciences, psychology, and other fields at Leiden University to advance the study of the evolution of human intelligence, behaviour, and societies. A dynamic programme of lectures, poster presentations, panel discussions, and social activities will take place across various locations

Compared to classical algorithms, which have enjoyed decades of development, many quantum algorithms are still in their infancy. The key technologies contained in classical algorithms often offer exponential advantages. This project will add these key technologies to quantum algorithms so that they enjoy the same advantage on top of the quantum acceleration. The resulting new line of quantum algorithms is therefore more able to compete with their classical counterparts. This closes an important gap on the road to a "quantum advantage," a first experiment showing that quantum algorithms can be faster than classical algorithms for a useful task.

Speech disorders impact tens of millions of people worldwide, hindering communication and significantly impacting daily lives. This project aims to develop a portable, real-time intelligent system tailored for speech disorder therapy. Combining advanced large language model technology with specific energy-efficient hardware design, the system provides real-time feedback and personalized guidance while maintaining low power consumption. Running entirely on mobile devices, the portable system ensures user privacy and therapy accessibility anytime, anywhere. This innovation makes speech therapy more accessible and effective, reduces clinical burdens, and helps individuals with speech disorders regain confidence and independence in communication.

Data mining (DM) is becoming ever more important as ever larger amounts and new forms of data are being generated in science and technology. Understanding the behavior of learning algorithms on various types of data is key in successfully applying them, and requires extensive experimentation. In this proposal, we aim to enable massively collaborative data mining research, through which the currently highly scattered data mining experiments from many researchers are collected and organized into a coherent, online experiment repository that can be accessed by anyone, in any form, from anywhere. We will develop easy-to-use tools that automatically evaluate data mining algorithms according to best practices in DM experimentation, explore novel types of meta-learning that measure their behavior under different conditions, and organize the ensuing results in a central, public database. This will concentrate the empirical knowledge of the field in one place, allow researchers to explore and build on each other?s results, make experimentation faster and more reliable, enhance insight into algorithm behavior, and disseminate novel findings more easily. Moreover, it facilitates collaboration with researchers in other sciences, and serves as an ideal educational platform. To illustrate the obtained benefits, we will exploit the resulting repository and novel meta-learning techniques to perform large-scale meta-learning studies that are nearly impossible today, on complex real-world bioinformatics data.

Models, i.e., simplified representations of real-world phenomena, were traditionally manually constructed based on knowledge and theory. In contrast, nowadays there is a strong trend towards automated, data-driven modelling, learning sophisticated models from large amounts of data. In this proposal I will argue that data science cannot always be fully automated and that experts often have relevant knowledge. Hence, there is a strong need for human-guided data science that integrates knowledge-driven and data-driven modelling. To this end I propose to develop methods for interactive model selection. Many data mining and machine learning tasks, such as classification, clustering, and outlier detection, can be formulated as model selection problems: given data and a set of models, the problem is to find the `optimal model. Currently no concrete, principled model selection methods exist that allow to integrate domain knowledge through interactive learning. To fill this gap in the literature, I propose iMDL, an innovative, interactive approach based on the minimum description length (MDL) principle. By iteratively eliciting feedback from an expert, a prior and constraints on the model class are learnt. This leads to accurate and meaningful models, while the MDL principle helps avoid overfitting and confirmation bias. Concrete iMDL instances will be developed for classification, clustering, and outlier detection. Methods will be evaluated 1) using simulations, to facilitate large-scale experiments; and 2) with case studies in the health care and aviation domains, to demonstrate the potential to interactively find accurate models that provide novel insights from data.