Seminar Series

Using a robot can present a situation that falls between using a tool, embodying an artificial body, and interacting with another person. I will present a series of our recent studies investigating how operating a robot characterized by different forms and levels of decision-making autonomy can influence cognition and behaviour, with a focus on sense of control and trust, as well as our preliminary EEG results.

The law assumes that healthy adults are generally responsible for their actions and have the ability to control their behaviour based on rational and moral principles. This contrasts with some recent neuroscientific accounts of action control as strongly biased. I present a short overview of laboratory experiments, in which we measured sense of agency, as participants performed actions while they were in either an emotionally neutral state, in a fearful or angry state. We found that fear or anger reduced the subjective sense of control over an action outcome, even though the objective causal link between action and outcome remained the same. This gap between the objective facts of agency and a reduced subjective experience of agency under emotional conditions has important implications for society and law. I further outline the neuroscientific evidence for such legal classifications of responsibility, focussing on how emotional states modulate voluntary motor control in the brain.

Loneliness is recognized as a major public health concern. The influential Evolutionary Theory of Loneliness (ETL) posits that social threat hypervigilance (STH) is a key mechanism underlying the long-term sequelae of chronic loneliness. However, evidence for STH in loneliness remains inconsistent. To address this, we conducted a series of studies examining the association between loneliness and STH using self-report, behavioural, ecological, neurophysiological, and neuroimaging methods. Our findings suggest that while STH may be evident in subjective reports, it does not consistently manifest across other levels of analysis in lonely individuals. Potential explanations for this discrepancy will be discussed.

Distinct concepts and categories produce separable patterns of brain activation detectable through imaging techniques such as fMRI—for example, seeing an object evokes a different neural response than a face. However, healthy aging reduces the selectivity of these activations in older adults, known as neural dedifferentiation. Activation patterns for categories become more alike, correlating with cognitive decline. I use computational modelling with deep-learning vision models to simulate localized damage in the ventral visual stream, exploring how different forms of damage could drive age-related neural dedifferentiation. I will discuss the effects of different types of damage in varying parts of the visual network and how this type of modelling could be used to explain fMRI data.

Interpersonal relationships are a fundamental aspect of being a social animal. Humans can effortlessly form rich impressions of others from minimal information, even before direct interaction. A key question is how this complex representation arises so easily, and why, for some individuals, sparse data can trigger psychological distress or disability. Traditional approaches have linked general learning models with psychometric data to address this question, but these methods risk oversimplification, leaving the mechanisms of social cognition underspecified. In this work, I present a computational approach to self-other information generalization that distinguishes between self-to-other, other-to-self information transfer, and observational learning. Using data from teenagers, adults, and patients with borderline personality disorder and psychosis, I demonstrate that self-other integration evolves with age and plays a critical role in social navigation. Disruptions in this process can contribute to psychiatric symptoms. I also discuss the utility of this computational framework for AI, experimental psychology, and computational psychiatry, and how it may intersect with other facets of social cognition such as metacognition and hierarchical reasoning.

Can we successfully ignore information we deem irrelevant or unreliable? And can we become aware of how such information can bias our choices? The ability to introspect and evaluate the factors influencing our decisions constitutes a key metacognitive function but little is known of the cognitive processes that underlie this ability. In this talk, I will present new studies where we measured participants’ ability to voluntarily ignore information that could bias their choices. We found that irrelevant or unreliable information continued to bias decisions, even when explicitly labelled as such. Participants were able to become aware of some of those biases but did not manage to compensate for them, showing a dissociation between metacognitive knowledge and metacognitive control.

Optimal stopping problems provide a framework for studying decision-making under uncertainty. We investigated deviations from normative strategies in human decision-making and examined the neural mechanisms underlying these processes using EEG. Forty participants completed a beads task under two levels of uncertainties and choice types (draw or urn choice) during EEG recording. A Bayesian ideal observer model and parametrised models were used to predict participant behaviour. Participants systematically deviated from Bayesian ideal observer predictions, demonstrating undersampling behaviour, particularly in the more uncertain condition. P300 amplitudes were modulated by both uncertainty and choice type, with larger amplitudes observed for urn choices compared to draw choices. Beta oscillatory activity further differentiated uncertainty conditions and choice types, highlighting distinct roles for oscillations in uncertainty processing and decision commitment. Additionally, beta power was associated with trialwise model-derived action values, emphasising the relevance of beta oscillations in capturing decision-making dynamics. These findings demonstrate that deviations from normative strategies are consistent across manipulations and provide new insights into the neural correlates of evidence-gathering and decision-making.

The fight against climate change requires institutional solutions such as taxes or subsidies. The most effective climate policies bear global benefits, but they come at a cost to local communities. This creates a social dilemma. Using a cross-national survey (N = 604), we show descriptively that even citizens who are concerned about the threat of climate change are reluctant to endorse local policies if they believe others do not support such policies. This is a particular challenge in democracies, where citizens may directly or indirectly vote on climate policies. We propose a theoretical model in which voters favour globally-efficient policies if they believe that adopting such a policy, even unilaterally, improves compliance with local climate goals. To test this model, we introduce a new climate policy dilemma game. In a large-scale incentivised experiment (N = 1,730), we show that a sizable minority of players prefer a globally efficient policy even though it is locally costly. This support for globally efficient policies is partly driven by the anticipated positive incentive effect of such a policy on collective contributions towards climate change mitigation. Our study provides a test of causal mechanisms underlying the support for ambitious climate policies and, more broadly, of the causal effects of democratic processes on legitimacy and policy compliance.

The talk introduces a modern machine learning method called Conformal Prediction (CP)––a new branch of statistical inference ideally suitable as a framework for uncertainty quantification. We shall present the method and its use for reliable prediction and decision-making.

Each sensory modality has its own primary and secondary (“high-order”) thalamic nuclei. While the primary thalamic nuclei are well understood to relay sensory information from the periphery to the cortex, the role of high-order sensory nuclei is elusive. One hypothesis has been that high-order nuclei may support feature-based attention. If this is true, one would also expect the activity in different nuclei to reflect the degree to which modalities are or are not behaviorally relevant in a task. We trained head-fixed mice to attend to one sensory modality while ignoring a second modality and simultaneously recorded from high-order somatosensory (PO) and visual thalamus (pulvinar, LP). Training could switch the modality that maximally activated a secondary thalamic nucleus. Movements do not account for this dramatic switch. Secondary nuclei appear to encode behaviorally relevant, reward-predicting stimuli regardless of stimulus modality. This may facilitate cortical plasticity during learning by activating apical dendrites in layer 1.

Parsimony is a desirable feature of economic models but almost all human behaviors are characterized by vast individual variation that appears to defy parsimony. How much parsimony do we need to give up to capture the fundamental aspects of a population’s distributional preferences and to maintain high predictive ability? Using a Bayesian nonparametric clustering method that makes the trade-off between parsimony and descriptive accuracy explicit, we show that three preference types—an inequality averse, an altruistic and a predominantly selfish type—capture the essence of behavioral heterogeneity. These types independently emerge in four different data sets and are strikingly stable over time. They predict out-of-sample behavior equally well as a model that permits all individuals to differ and substantially better than a representative agent model and a state-of-the-art machine learning algorithm. Thus, a parsimonious model with three stable types captures key characteristics of distributional preferences and has excellent predictive power.