Computational Psychiatry

May 29-June 5, 2027

Director: Ray Dolan

University College London, UK

Co-Director: Ulman Lindenberger

Max Planck Institute for Human Development, Berlin, Germany

Faculty:

Stephen Fleming, University College London, UK

Klaus-Robert Müller, Technical University Berlin, Germany

Natalie Shapira, Northeastern University, Boston, USA

Wen Li, University of Texas Health Science Center at Houston, USA

Yunzhe Liu, Beijing Normal University, China

Rani Moran, Queen Mary University of London, UK

Ray Dolan, University College London, UK

Ulman Lindenberger, Max Planck Institute for Human Development, Berlin, Germany

Computational psychiatry is now more than fifteen years old. Emerging from advances in computational and cognitive neuroscience, it promised a new mechanistic understanding of the development and expression of psychiatric disorders, and ultimately more effective approaches to prevention and treatment. How far has this been realised, and will recent developments in artificial intelligence (AI) overcome obstacles that have hindered progress?

The Advanced Course will take critical stock of the field, including major conceptual and empirical advances, current shortcomings, and most promising directions for the next decade. We will examine computational accounts of key domains of human cognition and their disruption in psychiatric illness, including reinforcement learning, metacognition, emotion, cognitive maps, and language. We will also consider both how AI provides powerful new tools for understanding human cognition and psychopathology, and how AI might inform existing ideas about computation, representation and intervention in psychiatry and cognitive neuroscience in general. In so doing, the Advanced Course will emphasise current controversies, unresolved questions and emerging opportunities.

Stephen Fleming

Metacognition and reality monitoring: from neural circuits to subjective experience

Metacognition enables organisms to self-evaluate their own cognitive processes. In humans, this capacity underpins adaptive learning, social coordination and the formation of self-beliefs about skill and ability. A core motif underlying multiple forms of metacognition is the ability to estimate performance in a self-directed frame of reference, expressed as propositional confidence about one’s own behaviour and mental processes. Convergent behavioural, computational, and neural evidence across rodents, non-human primates, and humans indicates that these self-evaluations arise from structured transformations of uncertainty. Dynamic evidence accumulation processes provide candidate building blocks for local confidence, whereas multimodal prefrontal brain areas integrate these signals to support abstract, global self-beliefs. This architecture supports metacognitive control at multiple timescales, from rapid error correction to strategic task arbitration and social communication. In my talk, I will explore how dysfunctional metacognition can lead to disorders of reality monitoring and insight, in which erroneous mental representations are endorsed with aberrant confidence. By integrating computational, cross-species and clinical perspectives, metacognitive neuroscience bridges neural circuit mechanisms and subjective experience.

Natalie Shapira

AI and the mind: drawing lessons from each other

AI is changing how we study human behaviour while also creating a new kind of intelligent object to investigate. These lectures will explore this two-way relationship between AI and the science of the mind.

I will begin with the question of what AI can do for psychiatry and psychotherapy. Modern AI systems can extract information from rich behavioural data that is difficult to measure with conventional clinical instruments, opening possibilities for automated assessment, diagnosis, monitoring, and the discovery of new clinical signals. I will illustrate these possibilities with examples from work on psychological and therapeutic data, including analyses of individual patients and therapeutic dyads, and the detection of subtle interpersonal and intrapersonal phenomena such as therapist self-disclosure or synchrony between therapist and client. These examples also reveal a fundamental challenge: many psychological constructs are difficult to operationalise computationally. I will discuss our work on humour in psychotherapy as a case study in how to approach this problem without first requiring a precise computational definition of the phenomenon itself.

We will then turn the question around: what can the science of the mind teach us about AI? Psychological concepts such as beliefs, perspective-taking, social reasoning, and theory of mind offer useful tools for characterising AI behaviour. I will discuss my research on theory of mind in LLMs and the controversial question of what it means for an artificial system to exhibit behaviour that resembles human mental-state reasoning. This perspective also has direct implications for AI safety. AI agents can be influenced through vulnerabilities in their social and psychological processing. I will discuss our work on Agents of Chaos as an example of how seemingly psychological properties of AI agents can become attack surfaces.

This leads to a broader research vision: Mechanistic Empathy. Can we create healthy, prosocial AI systems, and if so, how? What do such mechanisms look like? What should we look for within AI? How can we strengthen them? We can ask whether the computational mechanisms underlying socially meaningful behaviour can be identified, manipulated, and understood. Mechanistic interpretability provides tools for tracing such mechanisms inside neural networks; psychology provides theories and phenomena that can guide what mechanisms to look for and what questions to ask.

The lectures will conclude with a hands-on exploration of model mechanisms using causal tracing on open-weight LLMs. No prior technical knowledge is required. Participants are encouraged to bring a personal laptop.

Rani Moran

Understanding human learning through reinforcement learning: insights and challenges

Reinforcement learning (RL) provides a computational framework for understanding how agents learn from experience to guide future behaviour. At its core is a simple idea: discrepancies between expected and observed outcomes are used to update expectations and shape subsequent choices.

Over the past decades, RL has become an influential framework for studying human learning and decision-making, providing insights into both adaptive behaviour and systematic biases. In the lectures, I will introduce the basic principles of RL and then focus on challenges that arise when using it to understand human behaviour. One challenge stems, paradoxically, from the framework’s success: RL has been used to characterise an increasing range of processes and biases, including confirmation bias, perseveration, and credit assignment to task-irrelevant features or rejected options, making it increasingly difficult to distinguish their contributions to behaviour. I will illustrate this with an example of how perseveration can masquerade as asymmetric learning, and hence as apparent confirmation bias.

A second challenge concerns what RL parameters actually measure: because learning and choice unfold sequentially and are linked through nonlinear processes, fitted parameters need not straightforwardly reflect the underlying psychological quantities they are intended to capture. Finally, I will discuss how we have recently used RL to study learning in contemporary information environments, examining how people learn when feedback comes from biased, unreliable, or explicitly untruthful sources, and what this can reveal about the challenges of forming beliefs in the age of misinformation and disinformation.

Yunzhe Liu

Hippocampal ripples and neural replay: Inference and creativity

How does the brain use past experience to solve new problems, and how can we move beyond familiar ideas when they constrain our thinking? These lectures will explore hippocampal ripples and neural replay in the human brain, introducing the underlying neuroscience before examining how interactions between the hippocampus and prefrontal cortex support inference and creativity.

Part 1: Understanding ripples and replay: mechanisms and measurement

I will begin by introducing hippocampal ripples, brief bursts of high-frequency neural activity, and replay, the sequential reactivation of neural representations. I will explain how these phenomena are related and discuss evidence for their roles in memory and cognition during sleep, rest and active behaviour. We will then consider how to measure them in humans: how intracranial recordings identify ripples, and how neural decoding and sequence analysis reveal replay, including in magnetoencephalography (MEG). Throughout, I will highlight what these measurements can tell us about the content and timing of neural processing, alongside the challenges of linking them to specific cognitive functions. This will provide a foundation for understanding how the brain retrieves and recombines experience to construct internal models of the world.

Part 2: From inference to creativity: testing mechanisms with brain stimulation

I will then present two studies examining how hippocampal ripples relate to changes in prefrontal representations. In a LEGO-like inference task, ripple-aligned replay brings together familiar elements, supporting representations of new solutions in medial prefrontal cortex. In a divergent thinking task, more original ideas emerge when ripple-aligned ventromedial prefrontal activity is transiently reduced, accompanied by a shift away from established schemas towards more remote concepts. Together, these findings suggest that hippocampal–prefrontal interactions can support both constructing coherent internal models and loosening their constraints on thought. I will introduce ripple-triggered prefrontal stimulation as a way to test the causal contribution of prefrontal activity during these brief windows, including evidence that stimulation can increase originality. Finally, I will discuss how these findings might inform our understanding of inflexible thought patterns and suggest future directions for targeted interventions.

Wen Li

Sensory cortical mechanisms of anxiety and fear memory

In this session, I will introduce a sensory-centered framework for understanding fear and anxiety, moving beyond traditional limbic-centered models. I will then consider how sensory cortices actively contribute to threat perception, fear learning, memory formation, and psychopathology. Drawing on convergent evidence from behavioral experiments, electrophysiology, and neuroimaging, I will explore multiple sensory mechanisms that contribute to fear and anxiety, including intrinsic sensory cortical disinhibition, altered cortical threat representations, and experience-dependent plasticity. Finally, I will consider how targeted interventions, including sensory-feature-based perceptual training and non-invasive neuromodulation, test the outlined framework. In effect, I will show how modifying sensory cortical mechanisms influences the expression of fear and anxiety. Together, the session will highlight sensory cortices not merely as passive recipients of emotional modulation, but as active contributors to fear and anxiety and as promising targets for intervention.

Experimental design and methodological innovation in human neuroscience

This session examines how rigorous experimental design and well-chosen methods can advance scientific discovery. First, I will illustrate how classical paradigms, strengthened by targeted manipulations and multiwave, multitask, and multisample assessments, can uncover novel neural mechanisms and establish the reliability and generalizability of findings. Studies of general sensory processing and threat-specific sensory representations in anxiety demonstrate how relatively simple EEG or MEG recordings, paired with thoughtful experimental designs, can yield fundamental insights into brain function. In a second part, I will consider intensive multimodal approaches, including simultaneous EEG–fMRI–tACS and AI-based computational methods, to address two distinct questions: how neural oscillations interact with large-scale brain networks, and how sensory cortical activity contributes to fear memory formation. These examples will highlight how complementary neural measurements and causal perturbations can be integrated to investigate complex brain mechanisms, including those that underpin anxiety and pathological fear.