Please find the abstracts of the doctoral students in alphabetical order below:
Moana Beyer, MPI for Human Development
Exploring the Association of the Physical Environment with Brain Structure and Psychosis through Boruta
Previous research has consistently highlighted the detrimental impact of city living on mental health, particularly on psychosis and schizophrenia, as well as brain structure. Given the global shift toward urbanisation, it becomes increasingly important to pinpoint the specific urban features responsible for these associations to inform urban planning and public health efforts.
To this end, we employed feature selection methods to identify the most influential features of the residential environment in predicting psychotic experiences and grey matter volume. Urban features encompassed measures of the physical environment, such as air pollution, noise pollution, population density, and exposure to greenspace, and features of the social environment, such as crime rates and income deprivation. The analyses predominantly relied on Boruta, a wrapper built around a random forest classifier algorithm (Kursa et al., 2010).
Our results revealed a moderate predictive power for both psychotic experiences and brain structure. Individual-specific aspects of the participants’ social environments emerged as the most robust variables when predicting psychotic experiences. Additionally, the analysis revealed a non-monotonic, negative association of exposure to natural environments with psychosis. Preliminary findings indicate that exposure to natural and green surroundings, along with air pollution levels, predict grey matter volume in regions including the frontal and hippocampal areas.
In conclusion, spending more time in residential natural environments may not only mitigate the risk of psychosis but also exert changes in brain structure, offering actionable strategies in urban planning to enhance residents’ mental well-being.
Cormac Dickson, University College London
Mind wandering mechanisms
During waking hours our attention fluctuates between external (sensory) content and internally generated content. This ability to perceptually decouple from our environment and let our mind wander is highly prevalent throughout our daily lives, yet the computations which underlie its content are poorly understood. According to a resource-rational account, the occurrence of mind-wandering (MW) may be driven by relative changes in the opportunity cost of engaging with the external environment which also motivates periods of rest and sleep. In addition, one of the key computations identified during both sleep and rest in humans and rodents, neural replay, has recently been implicated with the occurrence of MW via both intracranial recordings and neuroimaging. Bursts of neural replay were found to coincide with self-reports of internally generated thought and activation of the default mode network, commonly implicated in MW, suggesting neural replay and MW may share a common resource. My current experimental work is focused on developing a paradigm that sheds light on whether facilitating or inhibiting mindwandering alters performance on tasks hypothesised to require neural replay. Foundational to this research programme is the development of tools to manipulate and measure momentary mindwandering in human participants.
Hadeel Haj-Ali, University College London
Emotional Habituation is Associated with Risk-taking Escalation
Risk-taking is a fundamental facet of human nature that has the potential to yield rewards and foster personal growth, as exemplified by the act of exploration. However, when extreme, it can lead to dire consequences. Anecdotally, extreme risk-taking often starts with small, seemingly harmless, actions that gradually snowball into more extreme behaviours. Here we used a computerized gambling task (Experiment 1) as well as a Virtual Reality task (Experiment 2) to present empirical evidence of the gradual escalation of risk-taking behaviours in a controlled experimental setting. We assessed participants' emotional responses using self-reports and measured autonomic arousal indexed by Galvanic Skin Response (GSR). We predicted that the emotional response would habituate with repetition and that the rate of emotional habituation will be positively linked to the rate of risk-taking escalation. First, we observed that risk-taking escalated with repetition. Second, we found that emotion self-reports and arousal levels associated with risk-taking diminished with repetition. Lastly, across individuals, faster emotional habituation was associated with a more rapid escalation of risk-taking. Together, these findings suggest a role for emotional habituation in risk escalation, informing the development of interventions to curb dangerous risk-taking and prevention programs for vulnerable populations.
Anna Hall, University College London
A computational model of subjective hedonic experience in health and mental illness
Anhedonia is a loss of pleasure in enjoyable activities. It is important from a clinical perspective but also because it poses the fundamental question of how the brain constructs reward and pleasure. We consider a novel mechanistic account of subjective hedonic experience which builds on RL theories of physiological drive to reconceptualise subjective hedonic experience in terms of progress towards individually meaningful goals. In an initial questionnaire-based study, we found that subjects’ ratings of pleasure evoked by a real-world event were best explained by the subjective progress implied by the event towards their current goal, followed by the subjective importance of that goal. Subsequently, we aim to formalise progress under a novel RL framework. We are developing a graph-based task in which subjects navigate a maze towards a goal state and intermittently rate their experienced pleasure to explore how subjective hedonic experience is related to progress towards varied goals in RL terms. I will present preliminary findings from this task design as well as potential future developments, and discuss the potential implications of this work on our understanding of anhedonia.
Samuel Hewitt, University College London
Fluctuations in effort-based decision-making over time
Computational modelling has revealed that effort-based decision-making is underlined by differences in sensitivity to reward and effort. These differences have usually been interpreted as stable characteristics which differentiate individuals from each other, but cross-sectional studies are not designed to test this assumption. I will present results from one of my PhD projects in which we designed a novel smartphone-based protocol to assess naturalistic, day-to-day fluctuations in motivation over 2 weeks in healthy adults (age 18-45 years). We found that fluctuations in subjective motivation positively predict how much effort participants will choose to exert on a given day and replicate this in two independent samples (discovery N=25; replication N=130). The vast majority of the variance in the relationships between subjective state and effort-based choices was due to how people varied over time (in idiosyncratic ways), compared with how they varied from each other (on average). Using computational modelling, we can further test the extent to which fluctuating psychological states modulate the subjective value of effort-based choices. In summary, the willingness to exert effort typically varies significantly from day-to-day in predictable ways. This finding questions the extent to which effort-based choices and their underlying components, captured at any given moment, represent stable personality characteristics.
Moritz Ketzer, Humboldt-Universität zu Berlin
Bridging the Gap: Constructing Directed Acyclic Graphs for Multilevel Models
Recent work in causal inference has unlocked the capability to estimate a variety of causal quantities in parametric and nonparametric structural equation models (SEM). A highly informative causal quantity is the interventional distribution, which refers to the probability distribution of outcomes under a specified intervention. To define causal estimands, communicate causal assumptions, and identify causal estimands, directed acyclic graphs (DAG's) have evolved to be an integral methodology in causal inference. While it is relatively straightforward to construct DAGs for basic SEMs, a gap remains: How do we construct DAGs for models with nested and hierarchical structures? Our research aims to bridge this gap by exploring the construction of DAGs for multilevel models, like multilevel regression and multilevel SEM. I will provide a short introduction to causal inference and multilevel modelling, and then show how DAGs represent crucial and intuitive tools for scientists to reason and communicate about the assumptions, structure and intricacies of multi-level models.
Agnieszka Kulesza, MPI for Human Development
The not-so-tiny blue spot: Dopaminergic and noradrenergic modulation of aging memory and synaptic plasticity
The core of scientific pursuit lies in understanding observed phenomena by uncovering their underlying causes. While human cognitive neuroscience has shown associations between certain neuromodulatory systems and behavior, causal relations have rarely been demonstrated. To more directly link brain and behavior, in two projects I investigate within-person changes in neuromodulatory systems and their effects on cognition. The main emphasis of this presentation will be on causal associations between these levels of analysis.
Specifically, I focus on the locus coeruleus (LC) - a small brainstem nucleus, responsible for releasing noradrenaline and dopamine, that play a pivotal role in hippocampal synaptic plasticity and memory formation. In the first project, age-related changes in the LC, assessed using structural MRI, will be used to predict different trajectories of memory aging. I will trace a link between within-person LC changes and repeated multimodal memory assessments collected over a decade. I will propose a model to test the lead-lag relationships across neural and cognitive measures. In the second project, experimentally induced changes in LC activity will allow me to establish a more causal relationship between the LC and learning. In particular, I will investigate if and how non-invasive electrical stimulation of the vagus nerve, a nerve transmitting peripheral signals to the brain, can modulate LC activity in humans.
Taken together, both approaches will offer insights into different ways of measuring the associations between the LC neuromodulatory system and memory. This will essentially contribute to a more comprehensive and mechanistic picture of how LC neuromodulation contributes to memory decline in old age.
Erwann Le Lannou, University College London
When in doubt use your priors: Risk and ambiguity aversion as rational probability weighting
Formal approaches to risky decisions making often do not allow for individuals to adapt their preferences based on subtle changes in the environement. Here we present two studies where we show that individuals have a tendency to adapt their aversion to uncertainty depending on nuanced factors in the decision set. First we show that in ambiguity. Most formal existing models, rooted in linear interpretations of ambiguity attitude, and are limited in their ability to explain interactions between the degree of missing information and where it fits on a probability scale. It is also the case that extant models of ambiguity do not of themselves make predictions regarding decisions under risk, wherein outcome probabilities are explicit. To address this we introduce a Bayesian-inspired model where ambiguity attitudes hinge on individual priors. Our approach seamlessly integrates decisions made under both risk and ambiguity, postulating all gambles inherently embody a degree of ambiguity. Crucially, our model forecasts a non-linear interplay between the amount of undisclosed information and the distribution's expected value. In a set of behavioural experiments, we find support for this prediction, showing predicted interaction effects between the degree of missing information and the expected value of the choice. In a second study, we look at sequential decisions - indeed most work in the field has focused on single shot decisions. Here we hypothesise that effects of partial outcomes following initial choices lead to changes in risk preference and further pushes us to consider risk as a more fluid metric.
Anika Löwe, MPI for Human Development
Insight in adults, children and machines
While working on a task, humans sometimes have an insight that leads to a sudden and drastic performance improvement. Insights are often seen as a unique aspect of human cognition that is related to creativity and meta-cognitive reasoning.
I take a learning perspective on the mysterious aha-moments by comparing learning dynamics in humans and regularised neural networks in a perceptual decision task that includes a hidden regularity to solve the task more efficiently.
I find that insight-like behaviour can occur in simple artificial neural networks, even when trained gradually with gradient descent algorithms, shedding light on the computational origins of insight.
Children, who perform worse than adults on most cognitive tasks requiring focussed attention, have insights about a hidden task regularity to the same extent as afults. This suggests that insight-like flexible strategy updating is a remarkable exception of the protracted decision function abilities.
Further, I investigate the role that sleep plays in insight. As information restructuring and memory consolidation occur during sleep, this cognitive state is an interesting candidate for insights. The physiological characteristics of sleep stage 1 (N1 sleep) in particular could foster the generation of novel ideas while maintaining logical inference capabilities. I am testing this by conducting the same insight task, interrupted by a 20-minute rest period for participants to nap while EEG data is acquired to identify sleep stages.
Rita Moura Alexandre, University College London
Behavioural markers of pathological cognitive decline
Recent insights suggest that memory transcends simple recall; it in-volves integrating discrete events crucial for interpreting present contexts and predicting future events. This is vital across the lifespan and relevant when studying neurodegenerative diseases affecting the hippocampus.
Mild Cognitive Impairment (MCI) is defined as a cognitive decline that is not severe enough to significantly disrupt daily life and may be indicative of early Alzheimer’s Disease (AD). It poses a diagnostic challenge, as its subtleties are often undetected by standard neuropsychological tests, particularly those assessing hippocampal-associated memory decline.
We designed a novel computational task to examine age-related differ-ences in functions that are vulnerable to neurodegeneration: associative memory, inference, and schema learning. To ensure broad applicability, we have refined our methodology to engage specific memory-dependent cognitive processes, thereby offering insight into the underlying neural mechanisms at work.
The task is administered remotely, along with a working memory con-trol task and specific questionnaires, to differentiate between typical age-ing and the early signs of MCI. We will discuss the task design and pre-liminary results. Additionally, we plan to extend our participant base to include individuals with amnestic MCI and those with subjective cognitive complaints to further validate our findings.
Our investigation into the broader role of memory in inference is a step towards a more nuanced understanding of cognitive ageing.
Zoya Mooraj, MPI for Human Development
Longitudinal Assessments of Neural and Behavioral Variability in Aging
Neural variability has been shown to be a reliable index of individual differences in cognition and age. In particular, cross-sectional research demonstrates that aging is related to diminished modulation of neural variability in response to increasing task demands (meta-variability), indicating that within-person neural variability may serve as a sensitive marker of cognitive flexibility. Mechanistically, dopamine is considered to contribute to neural variability. Dopamine (particularly the D2 dopaminergic system) is thought to enable the switch between states of stability and flexibility, maintaining information processing by tuning signal and network dynamics. Furthermore, there is a well-established link between dopamine, aging and cognition. However, the link between neural meta-variability and dopamine is little-studied, and has never been assessed longitudinally. In this talk, I will discuss the background linking dopamine to neural meta-variability. I will then outline plans for an upcoming project leveraging a large (n=181) longitudinal, multimodal dataset of adults aged 60 at baseline, including measures of D2 dopamine capacity and working memory task fMRI. The main aim will be to investigate the relation of 5-year changes in D2 dopamine capacity to changes in meta-variability and aspects of cognitive performance.
Jakub Onysk, University College London
Behavioural and linguistic manifestations of context manipulation in relation to semantic processing – from self-worth to suicide.
Suicide is a perplexing behaviour, going against a hard-wired instinct of self-preservation. To better understand it, I follow up on findings regarding strong associations between concepts related to death and self in suicidal individuals as measured with the Implicit Association Task (IAT). While such findings are predictive of suicidal ideation and attempts, an understanding of what is behind such associations are lacking.
An overarching hypothesis is that behavioural manifestations of implicit associations stem from semantic representations of concepts individual's brain encodes. Differently weighted semantic structures may influence ideations, where in the case of strong semantic similarities between self and death concepts, suicidal ideations may occur.
In my initial exploration on an online sample, I aim to probe and temporarily change individual’s semantic structures by manipulating context through participants’ text reading and writing. Each text is uniquely defined by valence (good or bad) and by perspective (self or other).
I use behavioural tasks (e.g., IAT), to measure the effect of such manipulation on associative strength between concepts. Initial findings (n=70) demonstrate that there is an average effect of first-person perspective in the story and, separately, an average effect of positive valence on the strength of associations between positive concepts and oneself as measured with IAT reaction times differences. A mirroring contextual effects analysis is further discussed in the Large Language Models (LLMs) framework.
These results suggest that an individual’s underlying semantic structure may be susceptible to change as a function of context, but requires further exploration and methodological improvements.
India Pinhorn, University College London
Intrinsic and Instrumental value: Domain-general principles determine choice across material, cognitive and visual domains.
What makes something rewarding? Do humans use the same principles and computations to determine how rewarding new shoes are, as they do to determine how rewarding reading a Shakespeare sonnet is or viewing Monet’s Water Lilies? Here, we test the hypothesis that the brain uses domain-general principles to calculate the value of ‘things’, regardless of whether these ‘things’ are material (e.g., an eraser), cognitive (e.g., a piece of trivia) or visual (e.g., a painting). Participants rated material, visual and cognitive items on different dimensions and also selected between pairs of items. Across each domain (material, cognitive, visual) the value of an item was predicted by the same two principal components: one which describes an item’s instrumental utility (i.e., how ‘useful’ it is) and one that describes its ‘intrinsic value’ (which encompasses beauty, interest and meaning). These factors predict how much participants like an item and whether they choose it. Strikingly, the weight participants put on these two dimensions when making a choice was associated with mental health. Those individuals who reported better mental health put more weight on the ‘intrinsic’ dimension than on the 'instrumental' dimension when selecting between items. It is possible that sensitivity to intrinsic rewards promotes mental health and/or that good mental health is required to appreciate, and engage with, intrinsic rewards.
Kira Pohlmann, MPI for Human Development
Deep learning-based prediction of environmental factors from 3D brain data
Given urban population growth and its known impact on mental health, it's crucial to study the effect of the physical living environment on the brain. This project aims to predict environmental features from 3D whole-brain MRI data using a deep neural classifier network. We use data from the extensive UK Biobank dataset to train the network. Environmental factors are encoded binary in terms of population density (rural vs. urban) (n ≈ 8806), as well as continuously in terms of variables describing the participants’ immediate physical environments (e.g., percentages of greenspace and water) (n ≈ 31,330). We train two different network architectures and compare their performance: Simple Fully Connected Network (SFCN), a network specifically designed for 3D brain data, and Resnet-3D, which allows training the state-of-the-art classifier Resnet on 3D data.
A focus of our work lies in the utilization of class activation maps, which serve as a tool for visualizing the specific regions within the input that the classifier identifies as relevant for predicting the outcome. In our case, this may help to identify brain regions relevant to the prediction of environmental features. The networks are trained on structural images as well as preprocessed grey and white matter images of the brain. This enables differentiation between white and grey matter in the regions highlighted by the class activation maps.
With this work, we propose a novel approach for MRI brain data analysis of large-scale datasets and make an effort to explore the capabilities of deep learning in environmental neuroscience.
Juan Vidal-Perez, University College London
Learning from disinformation
Reinforcement Learning studies how animals and humans adapt to their environment by learning from experience which actions maximize reward acquisition. Extant research has focused on situations where learners receive veridical reward-feedback. However, as social beings, much of the feedback we receive is provided by others. Critically, such feedback may be misleading because others might manipulate us based on their own – even if well-intended – interests. A pressing question arises regarding how individuals adapt their behavior when confronted with potentially misleading feedback.
To address this question, we designed modified versions of the two-armed bandit task that informed participants about the outcomes of their choices through 'feedback agents’. Importantly the feedback provided by these agents may deviate from the true outcomes of the bandits. With this paradigm, we assessed how people adapt their learning when presented with different types of misinformative feedback, such as lies or biased information. Our results suggest that humans employ various strategies to filter out misinformation, depending on whether it is structured or stochastic. However, this filtering process is not foolproof and can lead to biases in our perception of the world. These insights can help us understand how learning from disinformation may contribute to deficits in mental health and social communication, such as political radicalization and belief in conspiracy narratives.
Xin Zhang, University College London
Fast neural computations for neurological phenomena
Neurological phenomena may be viewed as human experiences corresponding to a focal structural abnormality in the brain. For people with early neurodegenerative dementias such as Alzheimer’s disease, alterations in their experiences may not correspond to an obvious lesion on structural imaging but reflect subtle and discrete neuronal circuit-level changes associated with proteinopathies and neuronal degeneration. These circuit-level changes are difficult to detect using conventional clinical tools. Functional neuroimaging using magnetoencephalography (MEG) and electroencephalography (EEG) can characterise brain activity at high temporal resolution. They may be more suited to studying circuit-level changes and components of human behaviour. Neural replay is the sequential reactivation of previously experienced representational states at a compressed timescale (milliseconds), analogous to hippocampal replay in rodents. The study of neural replay via MEG/EEG may offer sufficient granularity to shed light on neurological phenomena and circuit-level changes in patients with early neurodegeneration. I will present a brief case study of a patient with a neurological phenomenon and then introduce our current study investigating neural replay in early Alzheimer’s disease.