Abstracts

Nathaniel Daw, Princeton University
Computational mechanisms of decision making: Individual and circumstantial differences
Plenary Keynote Lecture
A long line of research in human and animal behavioural neuroscience suggests that the brain has multiple, competing systems supporting decision making: more goal-directed and deliberative vs. more habitual or automatic. Recent work aims to clarify this distinction computationally, and permit distinguishing these systems experimentally, by associating these systems with different families of algorithms, known as model-based and model-free reinforcement learning. This perspective may help to understand how the brain arbitrates between these systems, via the differential costs and benefits of the two approaches. I review experimental and theoretical work that relies on this approach to shed light both on the circumstances (within an individual) when either system is deployed, and (across individuals) how differential reliance on either system contributes to patterns of individual differences in decision making. For instance, the reliance on model-based learning depends on the availability of cognitive resources such as working memory, and waxes and wanes with development and normal ageing. It also is affected by neurological disorders such as Parkinson’s disease. Moreover, this line of research supports the longstanding suggestion that an imbalance between these systems might underlie the compulsive nature of pathological conditions such as drug abuse; indeed, an under-reliance on model-based decision making specifically tracks symptoms of compulsive  psychopathology but not other pathologies such as depressive symptoms.

Ben Eppinger, Concordia University
Adult age differences in learning and decision making: From decisions from description to complex state spaces
Plenary Teaching Lecture
In the first part of this lecture I will provide a brief overview on adult age differences in decisions from description and experience. I will then show some recent findings on age-related changes in learning mechanisms with a focus on the learning of more complex task structures and state spaces. I will outline several problems with the current experimental designs and analysis approaches and will show how computational methods may provide one way out of these dilemmas. In the second part I will highlight a few questions for future research that I find exciting. I will particularly focus on age differences in learning in dynamically changing environments, on the question how we prioritize information in multidimensional state spaces, and how we arbitrate between different learning and decision strategies.

Douglas D. Garrett, Max Planck UCL Centre for Computational Psychiatry and Ageing Research, Berlin
Presentation of research group
Within the Lifespan Neural Dynamics Group (LNDG) in the Max Planck UCL Centre, we seek to understand the cognitive, developmental, and neurobiological basis for why human brain activity fluctuates so markedly from moment-to-moment. Typically, researchers conceive of signal variability as neural “noise,” a nuisance factor that presumably interferes with the efficiency of neural processes. However, increasing evidence supports that some degree of moment-to-moment signal variability is needed for neural communication, flexibility, and adaptability. The LNDG has focused on the manner in which brain signal variability reflects broad-scale cognition and ageing. In line with extant computational and animal research on the benefits of optimal neural variability, our work indicates that a possible basis for cognitive deficits associated with normal ageing may be a critical lack of moment-to-moment variability in older brains. Overall, our findings converge upon an initial characterisation of relations among age, cognitive performance, dopamine, and brain signal variability from which ageing could be conceived of as a gross “dedifferentiation” of brain signal dynamics. Recently funded by a five-year Emmy Nöther award from the German Research Foundation (DFG), our group will continue to investigate individual differences and longitudinal change in three key factors (brain structure, static/dynamic functional connectivity, and dopamine) that may yield joint reductions in brain signal variability (fMRI, EEG) and cognition with age. Requiring a multidisciplinary, multi-method approach, our research goals are designed to propel brain signal variability forward as a principled method for understanding the ageing brain.

Rogier Kievit, Cambridge University
Structural equation modelling of brain and behaviour
Plenary Teaching Lecture
Arguably the single most fundamental question in cognitive neuroscience is how to conceptualise the relationship between brain and behaviour. I will propose that SEM is uniquely powerful in translating theories about brain and mind into hypotheses that can be tested in neurocognitive datasets. First, I will lay the theoretical foundations, proposing that contrasting theoretical perspectives on this question can be tested empirically as SEM, and argue that supervenience, characterised by a many-to-one mapping between brain and behaviour is best supported by the data. Next, I will illustrate the power of SEM in testing theories of neurocognitive ageing. In the Cam-CAN cohort, I will propose a watershed model of individual differences in fluid reasoning, combining white matter integrity, processing speed and fluid reasoning in a principled manner in a single statistical model. Using equality constrained multigroup models I will examine and discuss some evidence for age differentiation within, but not between, neural (white and grey matter) and cognitive domains. In the final section of the talk I will describe recent applications in longitudinal SEM. I will show how bivariate latent change score models can provide evidence for mutualism during adolescence and early adulthood (NSPN, N=776), such that vocabulary and reasoning show mutually beneficial interactions, and illustrate how latent growth modelling can be used to detect and predict sleep stages in MEG resting state analyses (Cam-CAN, N=640). I will conclude by discussing limitations and future directions of SEM as a uniquely powerful tool in neurocognitive ageing.

Rani Moran, Max Planck UCL Centre for Computational Psychiatry and Ageing Research, London
Beyond basic perceptual decisions: Temporal and spatial uncertainty, biases, heterogeneity and post-choice integration
Plenary Teaching Lecture
In typical perceptual decision-making paradigms participants observe a stimulus and are asked to decide as quickly and as accurately as possible between one of two responses. Models of perceptual decisions attempt to provide an account for performance in terms of accuracy and choice-RT distributions. The success of a model to account for these dependent variables is considered to both validate and quantify the putative choice-underlying cognitive mechanisms. The central framework in the research of perceptual decision making, the theory of sequential sampling, suggests that choices are driven by accumulation-to-threshold mechanisms. In most basic task-setups, observers benefit from temporal and spatial certainty with respect to the timing and location of the decisionstimulus, i.e., stimuli appear in a well-defined and an expected time and location and so participants know exactly when and where they need to integrate the relevant perceptual information. Additionally, different trials are similar in difficulty and both response options are apriori equally likely to be correct. Finally, observers are only asked to make a choice and no further judgment is required. In recent years I have been studying cognitively-challenging tasks that relax these ‘basic conditions’, thus placing observers in more demanding choice-situations. These tasks were designed to test how observers detect signals under temporal and spatial uncertainty, how they make optimal decisions in heterogeneous and biased environments and what underlies their confidence in the correctness of their choices. Accounting for human performance in these tasks required extending standard sequential sampling models with additional strategic and dynamic components. These extensions convey new insights with respect to the important and multifaceted role that is served by flexible top-down control processes in complex decision situations.

Janaina Mourao-Miranda, Max Planck UCL Centre for Computational Psychiatry and Ageing Research, London
Machine learning as a tool to investigate brain-behaviour relationships
Plenary Keynote Lecture
Neuroimaging and machine learning techniques show potential as tools to identify biological measures that may help diagnosis and prognosis of mental health disorders. So far, however, most of the studies using these techniques have focused on binary classification problems, i.e. they summarize the clinical assessment into a single measure and the output of the models is limited to a probability value and in most cases a binary decision (patients/healthy control). Although these studies represent an important advance in the field, they provide limited information about the underlying biological mechanisms of the diseases and do not enable patient stratification. Considering the complexity of mental health disorders, it is potentially beneficial to embed a multidimensional description of the disorder into the models. In this talk I will present the machine learning framework commonly used for analyzing neuroimaging data and discuss some limitations of this framework. I will then show examples of how machine learning techniques can be applied to investigate relationships between multivariate measures of brain anatomy or function and multidimensional descriptions of mental health disorders.

Tim Pleskac, Max Planck Institute for Human Development
The dynamic nature of confidence
Plenary Teaching Lecture
How confident are you in your choice? Such a simple question for people to answer. Yet, modelling how people answer that question has proven challenging. Part of the challenge has been that it has been unclear what information is used to make a choice and what information is used to make a confident judgment. Another part of the challenge is that confidence is conceptualised as a static variable that does not change over time. In this talk, I will review my work that addresses both of these challenges. I will show that choice and confidence are based on the same evidence, but in contrast to recent neuro-computational models of confidence this evidence does not conform to what would be expected if the evidence reflected the likelihood of the data for the given hypotheses (i.e., the evidence accumulation process does not conform to a Bayesian optimal process). Instead for both choices and confidence people over-emphasize the strength of the information relative to the weight. I will also show that the confidence people express in their choice is not a simple snapshot of the accumulated evidence at the time of a choice. Instead confidence changes and it can change quickly on the order of milliseconds. Part of its dynamic nature is due to the contribution of post decisional processing of evidence. But, the dynamics are also due to confidence reflecting other aspects including the feedback people are given, new incoming information, and even the very act of making a choice. Together these results support an overall hypothesis that choice and confidence are the product of the same evidence accumulation process. Understanding this interrelationship can help us not only understand the cognitive and neural processes of evidence accumulation, but also understand how and why and when people are accurate and inaccurate in their choices and the judgments they make in the world.

Robb Rutledge, Max Planck UCL Centre for Computational Psychiatry and Ageing Research, London
Presentation of research group
Depression is the leading cause of disability worldwide, affecting more than 300 million people. Unfortunately, current antidepressant treatments do not help many of those who suffer from depression. It is now accepted that depression can result from a variety of different sources, much like a cough can have many different underlying causes. There is at present no reliable way for a psychiatrist to know which treatment is likely to be most effective for helping a particular depressed individual. Furthermore, researchers have not yet managed to provide a clear picture of what determines if, and when, an individual’s mood will worsen and what happens in the brain when mood changes. This lack of understanding of the determinants of mood also makes it difficult to develop new treatments for mood disorders like depression. My recent research has shown that it is possible to measure momentary subjective states like happiness. We can predict precisely how happiness will change from moment to moment during a decision-making game played on smartphones by over 18,000 players worldwide, and happiness can be manipulated using dopaminergic drugs and related to neural activity in dopamine projection areas measured with functional neuroimaging. Supported by a Career Development Award from the UK Medical Research Council, my research group will quantify how mood is determined across a variety of decision
environments and then ask whether this improved understanding of mood can be used to better understand depression. This research will examine the neural circuits that determine mood in both healthy and depressed individuals. We will use a combination of laboratory tasks and longitudinal testing with smartphones to track previously depressed individuals and to ask whether certain trajectories in the model parameters that capture both mood dynamics and decision processes are predictive of relapse.

Nico Schuck, Princeton University
Age differences in neural representations during spatial navigation
Plenary Teaching Lecture
The detailed representations that underlie human memory and decision making often remain occluded from direct observation. I will present a recent study in which we used fMRI to understand the content of neural representations that underlie spatial navigation. While volumetric and univariate fMRI analyses of a spatial navigation task offered only broader insights into age-related performance differences, model-based analyses of brain activity and behaviour suggested qualitatively different spatial representations in younger versus older adults. Overall, these results suggest ageing may be accompanied by changes in the content of relevant neural representations during spatial navigation.

Christopher Summerfield, University of Oxford
Optimality and suboptimality in human decision making
Plenary Keynote Lecture
Humans make "near-optimal" category judgments about noisy sensory stimuli, but on cognitive tasks they often exhibit systematic biases that fail to maximise economic outcomes. In my talk, I will discuss why. I will argue that because the ideal observer framework considers only noise that arises during sensory encoding, it frequently misspecifies the decision policy that will maximise rewards. When we also consider "late" noise - that arising during information integration - cognitive biases can often be reframed as efficient, reward-maximising policies. I will discuss with reference to data and modelling from tasks involving perceptual averaging, transitive choices and decoy effects in multi-alternative choices.

Manuel Völkle, Humboldt-Universität zu Berlin
Continuous time modelling: On the role of time in the search for mechanisms of the human mind, brain, and behaviour
Plenary Keynote Lecture
Understanding the mechanisms of the human mind, brain, and behaviour lies at the heart of psychological research. In this presentation, I want to argue that the time required for psychological processes to unfold, may be key to a better understanding of the underlying causal mechanisms. Unfortunately, however, the special role of time is often not given sufficient consideration in the design and analysis of longitudinal studies. After distinguishing between static and dynamic models of change, I will discuss problems in the analysis of dynamic processes related to unequal time intervals and will introduce continuous time modelling as a way to overcome these problems. Continuous time models make use of stochastic differential equations, whose parameters are estimated based on discrete time observations. As such, they not only represent a more efficient way of analysing existing data, but may also help in improving the design of studies, as well as our understanding of the data generating processes. To illustrate the latter two points, I reconsider the meaning of missing values and mediation processes in longitudinal studies from a continuous time perspective. In the remainder of the talk, I will focus on ongoing research regarding causality, unobserved heterogeneity, hierarchical models, current software development, and related statistical approaches. I will end with an outlook on the future of studying the mechanisms
underlying psychological processes and how a change in perspective from discrete time to continuous time modelling may be instrumental in this regard.

Go to Editor View