Summer School Impressions

The summer school opened with the foundations of decision-making research: utility theory, reinforcement learning (RL), and Signal Detection Theory (SDT).

Gabriele Bellucci introduced utility theory, showing how we assign value to options and how RL allows us to learn these values over experience. He then extended the discussion to social decision-making, presenting fairness models (Fehr & Schmidt, 1999) and Bayesian approaches to mentalizing. Keiji Ota complemented this with a hands-on session on risky decision-making, where we explored certainty equivalents, expected utility, and probability weighting.

Later, Carolina Feher da Silva contrasted model-based (goal-directed, effortful) and model-free (habitual, automatic) learning, presenting her Magic Carpet task (Feher da Silva & Hare, 2020) and sparking discussion about reproducibility in RL-based research. Finally, Matteo Lisi showed how confidence bias and noise shape deviations from optimal behavior. His SDT tutorial introduced d′ as a measure of evidence strength and showed how ROC curves can capture sensitivity in decision-making.

A first day that gave us the essential building blocks for thinking about decisions, from individual choice to social interaction.

Day 2 – Machine Learning

The second day turned toward artificial intelligence and its ethical, mathematical, and computational underpinnings.

Chris Watkins provided a historical overview of AI and large language models (LLMs), explaining how tokenization, BERT, and transformers underpin GPT-style models. He also raised ethical concerns about their training. Agnieszka Mensfelt built on this, discussing Agentic AI and the central challenge of value alignment in systems where no universal value system exists.

Robin Schiewer then dug into the modeling basis of RL through Markov Decision Processes (POMDPs). He emphasized the exploration–exploitation trade-off and highlighted how world-model simulations allow safer training, such as for autonomous vehicles. To close, David Kappel tackled uncertainty in deep neural networks. Since exact Bayesian learning is often intractable, he showed how variational Bayes with KL divergence offers a tractable approximation, making uncertainty estimation feasible in deep networks.

A thought-provoking day that balanced the ethical stakes of AI with the mathematical foundations of modern machine learning.

Day 3 – Social Sciences

Day 3 demonstrated how computational tools can reveal insights into social and political behavior.

Sílvia Majó-Vázquez addressed polarization and fragmentation in democratic states, using graph theory to uncover patterns at both aggregated and individual levels. Next, Andreu Casas introduced supervised classifiers for text, images, and multimodal data. Moving from n-gram models to BERT-like architectures, he demonstrated their relevance for social science applications, such as hate speech detection and protest material analysis, supported by a hands-on coding session using HuggingFace’s packages.

Finally, Chris Hanretty applied these methods to parliamentary speech analysis, showing that older politicians tend to focus less on future-oriented topics. His work highlighted how machine classification and fine-tuning can reveal behavioral patterns in massive text corpora.

A day that showcased the value of computational methods in making sense of political communication and democratic processes.

Day 4 (Pt 1) – Computational Psychiatry

The first part of Day 4 explored how computational models can illuminate mental health and psychiatry.

Jen Dykxhoorn discussed the social determinants of mental health, showing how factors at the individual, family, community, and structural levels affect outcomes. Using PCA, she found that higher social exclusion – especially among migrants and refugees – was linked to poorer mental health.

Francesco Scaramozzino then presented computational accounts of psychosis, framing it as inference under uncertainty. He explained the Predictive Processing Account and the free-energy principle, showing how minimizing unpredictability in internal states can help us understand psychotic-like experiences.

Alex Pike demonstrated how reinforcement learning models can clarify mental health mechanisms, supported by meta-analysis evidence. She highlighted the importance of parameter recovery and introduced Bayesian Model Averaging as a way to strengthen modeling conclusions.

A day which showed how computational methods can help us deepen our understanding of mental health.

Day 4 (Pt 2) – Applied Research and Inclusion

The second half of Day 4 highlighted scalable research methods and equity in research.

Kelly Donegan discussed moving science beyond the lab using online and mobile cognitive tasks. By gamifying assessments and making them accessible digitally, tools like Neuroka (which she developed during her PhD) enable scalable mental health research. Then, Alex Lloyd highlighted Patient and Public Involvement (PPI), advocating for collaboration with public and lived-experience experts at all stages of research, particularly in youth mental health studies. An Equity, Diversity, and Inclusion (EDI) panel chaired by Nura Sidarus featured Vanessa Pinfold, Romana Saleh, and Tiarna Lee. Panelists shared personal experiences emphasizing the importance of inclusive and supportive research environments, addressing challenges faced by marginalized groups, hidden disabilities, and the integration of charities into research processes.

This session underscored that impactful mental health research requires not only robust models but also inclusive practices and public engagement.

Day 5 – Genetics, Social Learning and Cultural Evolution

The final day connected genetics and social transmission with computational models of cultural evolution.

Fabio Manfredini presented his work linking RNA production in honeybees to the waggle dance through PCA. He explained how machine learning can integrate across data layers to improve prediction.

Gabriele Bellucci then turned to cultural evolution, presenting models of social transmission that demonstrate how group size affects the balance between cultural loss and innovation (Derex et al. 2013, Toyokawa et al. 2019).

Will Hoppitt introduced Network-Based Diffusion Analysis (NBDA), a method to detect and quantify social learning via the s parameter, while also mapping typical pathways of information transfer. Michael Chimento closed the week with a tutorial putting NBDA into practice.

A closing day that tied together biology, culture, and computational modeling, reminding us of the many levels at which decision sciences operate.