P001 - A Kalman filter extension for hybrid models of the two-step sequential decision task.
Biologische Psychologie, Universität zu Köln
Introduction: Reinforcement Learning (RL) tasks map onto a wide range of real-world scenarios and play a central role in decision-making research. One prominent example is the two-step task (a sequential Markov decision task; Daw et al., 2011) which is widely employed to study model-based (MB) and model-free (MF) influences on human choice behavior. While findings from other paradigms with fluctuating reinforcement schemes have stressed the influence of uncertainty (e.g. Daw et al., 2006, Chakroun et al., 2020), computational accounts of behavior on the two-step task thus far lack an incorporation of uncertainty. Methods: Here we extended the standard hybrid model of MB and MF control (Daw et al., 2011; Otto et al., 2013) by incorporating uncertainty-based updating via the Kalman-Filter (Kalman, 1960). Further model variants included parameters for directed exploration as well as perseveration. Models were evaluated using data from n=39 healthy participants who completed 300 trials of the two-step task. Results: The Kalman Filter extension with additional parameters for directed exploration and perseveration outperformed standard hybrid models. Model parameters exhibited associations with corresponding indices from traditional model-free analyses. Discussion: Our results converge with recent criticism regarding a simplistic demarcation of two competing learning systems (MB vs. MF) and instead stress the central role of uncertainty in guiding human choices in volatile environments. Results extend findings from other paradigms (Daw et al., 2006, Chakroun et al., 2020) and reveal evidence for directed exploration in two-step task behavior. Conceptual and methodological implications are discussed.
P002 - A preregistered proof-of-concept study of motor imagery-based fMRI neurofeedback training in stroke survivors
University of Münster
Introduction: Motor imagery-based functional magnetic resonance imaging neurofeedback (fMRI-NF) training for potential new therapeutic techniques that aim to improve motor impairment. In this proof-of-concept study, we translated an fMRI neurofeedback paradigm previously studied in healthy participants to stroke survivors using Open Science principles.
Methods: Real-time fMRI analyses were completed in TurboBrain Voyager. Visual feedback was provided from the supplementary motor area (SMA) targeting two different neurofeedback target levels (low and high) and presented to patients via customised Python scripts. The study introduced a Bayesian sequential sampling plan, which allows 1) flexible stopping, 2) providing evidence for a null effect, and 3) incorporating prior knowledge to yield higher sensitivity. The sampling plan, a priori hypotheses, and all planned analysis were preregistered to mitigate potential publication/researcher biases, all data and code were made available to allow for reproducibility (https://osf.io/y69jc/).
Results: At the group level, we found only anecdotal evidence for the preregistered hypotheses. At the individual level, we found anecdotal to moderate evidence for the absence of the hypothesized graded effect for most subjects.
Discussion: The presented null findings are relevant for future attempts to employ fMRI-NF training in stroke survivors. Unforeseen difficulties in the translation of our paradigm to a clinical setting required well documented deviations from the preregistered protocol. Taken together, this work provides new insights about the feasibility of motor imagery-based graded fMRI-NF training in MCA stroke survivors and it can serve as a template for a comprehensive study preregistration of a complex neuroimaging experiment in a clinical population.
P003 - A workflow for open and reproducible fMRI studies
1Max Planck Research Group NeuroCode, Max Planck Institute for Human Development, Deutschland; 2Max Planck UCL Centre for Computational Psychiatry and Ageing Research, Berlin, Germany
Achieving computational reproducibility and accessible data sharing can be challenging, in particular for neuroimaging research that involves large amounts of heterogeneous data and code. Here, we showcase a workflow that combines several software tools to allow reproducibility and transparent sharing of code and data of a human fMRI study. We recently published an open-access paper (Wittkuhn & Schuck, 2021, Nature Communications) together with the code, data and computational environments needed to reproduce the reported results. We shared > 10 datasets via GIN (G-Node Infrastructure) as modular version-controlled units, including fMRI data organized in BIDS format and derived data, such as pre-processed fMRI data and data quality metrics. Research data was version-controlled using DataLad. Following the DataLad YODA principles, we nested datasets as modular units, allowing to better establish data provenance, i.e., a clear overview which code used which input data to produce which output data. Code that reproduced the analyses was integrated with additional documentation using RMarkdown notebooks. The notebooks were automatically executed using continuous integration on GitLab. In this process, data was retrieved from GIN using DataLad, the notebooks were rendered and deployed to a website (https://wittkuhn.mpib.berlin/highspeed/). Code execution was performed using software containers (Docker and Singularity) and virtual environments, allowing to reproduce the computational environment. We will discuss ongoing improvements of this approach, including the combination of software containers, code recipes using Makefiles and the role of documentation.
P004 - Automating the Construction of Scientific Models to Explain Human Information Processing
Princeton University, United States of America
Various empirical sciences, including psychology and neuroscience, are in the midst of a replicability crisis. This crisis is fueled by limited temporal and monetary resources to test and integrate an increasingly large number of theories and experimental phenomena, as well as a lacking standardization of scientific methods. We seek to overcome these limitations by integrating existing machine learning techniques into a closed-loop system for the generation, estimation and validation of interpretable scientific models.
We introduce and evaluate a novel method for recovering quantitative models of human information processing using differentiable architecture search (DARTS)—a technique that led to breakthroughs in the automated construction and parameterization of models in machine learning (e.g. computer vision) but that has not yet been applied to the discovery of models of brain function. This method treats scientific models as computation graphs, and leverages automatic differentiation to derive such models from empirical data. We evaluate the performance of this method based on its ability to recover three quantitative models of human information processing from synthetic data.
Our results indicate that this method is capable of recovering basic quantitative motifs from models of psychophysics and decision making. We also identify weaknesses of this method in recovering models of exponential learning.
Findings of this study highlight the utility of DARTS for the automated discovery of quantitative models of brain function. We invite interested researchers to evaluate this method based on other scientific models, and provide open access to a documented implementation of our evaluation pipeline.
P005 - Dynamics of fMRI patterns reflect sub-second activation sequences and reveal replay in human visual cortex
1Max Planck Research Group NeuroCode, Max Planck Institute for Human Development, Berlin, Germany; 2Max Planck UCL Centre for Computational Psychiatry and Ageing Research, Berlin, Germany
Neural computations are often fast and anatomically localized. Yet, investigating such computations in humans is challenging because non-invasive methods have either high temporal or spatial resolution, but not both. Of particular relevance, fast neural replay is known to occur throughout the brain in a coordinated fashion about which little is known. We develop a multivariate analysis method for functional magnetic resonance imaging that makes it possible to study sequentially activated neural patterns separated by less than 100 ms with precise spatial resolution. Human participants viewed five images individually and sequentially with speeds up to 32 ms between items. Probabilistic pattern classifiers were trained on activation patterns in visual and ventrotemporal cortex during individual image trials. Applied to sequence trials, probabilistic classifier time courses allow the detection of neural representations and their order. Order detection remains possible at speeds up to 32 ms between items (plus 100 ms per item). The frequency spectrum of the sequentiality metric distinguishes between sub- versus supra-second sequences. Importantly, applied to resting-state data our method reveals fast replay of task-related stimuli in visual cortex. This indicates that non-hippocampal replay occurs even after tasks without memory requirements and shows that our method can be used to detect such spontaneously occurring replay.
P006 - Effects of single-session transcranial direct current stimulation on reactive response inhibition
1University College Dublin, School of Psychology, Irland, Republik; 2Trier University, Department of Cognitive Psychology, Germany; 3Max Planck Institute for Human Cognitive and Brain Sciences, Research Group Cognition and Plasticity, Germany
Transcranial direct current stimulation (tDCS) is widely used to explore the role of various cortical regions for different cognitive processes. Reactive response inhibition refers to the process of stopping an already initiated response, which is crucial for efficient everyday performance. In recent years, tDCS studies reported polarity- and time-dependent effects on response inhibition. Given the rapid increase in tDCS application across disciplines, it is crucial to systematically explore the existing tDCS literature to increase the current understanding of potential modulatory effects and limitations of different approaches. Consequently, we performed a systematic review according to PRISMA guidelines on the modulatory effects of tDCS on response inhibition as measured by the Stop-Signal Task, a standard measure for response inhibition. The final dataset includes 31 studies which show a large variation in methodology, resulting in heterogenous effects of tDCS on task performance. Further, methodological reporting procedures and data availability vary drastically which makes replication of studies or confirmation of results partially impossible in up to 2/3 of studies. As a main finding, results show that anodal tDCS over the right prefrontal cortex has the potential to enhance response inhibition when applied before a performance measurement. We note that partially sub-optimal choices in study design and methodology as well as lacking consistency in reporting procedures may impede valid conclusions and may have obscured the effects of tDCS on response inhibition in some previous studies. Finally, we outline future directions to improve tDCS research in studies of cognition in general and response inhibition in particular.
P007 - Effects of transcranial alternating current stimulation on spontaneous, transient brain states – A Hidden Markov Model approach
University of Oldenburg, Deutschland
Non-invasive techniques to electrically stimulate the brain are increasingly used in human neuroscience and offer new avenues to treat brain disorders. However, their often weak and variable effects have raised concerns in the scientific community. A possible factor influencing the efficacy of these methods is the dependency on brain-states. This state-dependency is usually investigated by experimentally inducing states for long periods of time, which is in stark contrast to the timescales the brain usually operates at. Here, we utilized a novel Hidden Markov Model (HMM) framework to decompose magnetoencephalography (MEG) data of participants who received 20-min of transcranial alternating current stimulation (tACS) at alpha frequency or sham stimulation into transient brain-states with distinct spatial, spectral and connectivity profiles. We compared the change in power from a 10-min baseline period before stimulation and a post-stimulation period immediately after tACS. We found that only one out of the four spontaneous brain-states was susceptible to tACS. No or only marginal effects were found in the remaining states. TACS did not influence the time spent in each state. Our results suggest that effects of tACS may be mediated by a hidden, spontaneous state-dependency and provide novel insights to the changes in oscillatory activity underlying effects of tACS.
P008 - Exploration behavior in recurrent neural networks during reinforcement learning in volatile environments.
1Biologische Psychologie, Universität zu Köln, Deutschland; 2University of California, Berkley, USA
Introduction: Recurrent neural networks (RNNs) are a promising model of human cognition (Gershman & Ölveczky, 2020; Botvinick et al., 2020). Past research showed that RNNs with computation noise show resilience to adverse conditions in reinforcement learning (Findling & Wyart 2020). The multi-armed restless bandit task is a non-stationary reinforcement learning problem, thus options must be continuously explored to increase cumulative reward (Daw et al., 2006). Here, human subjects show evidence for both random and directed (uncertainty-based) exploration (Chakroun et al., 2020). We investigated the performance of noisy vs. non-noisy RNNs in restless bandit problems and compare it to human performance.
Methods: We trained RNNs (48 units, Findling et al. 2020) using the REINFORCE algorithm on binary four-armed restless bandit problems with fixed training volatility, and examined their performance on the same task structure across a range of test volatilities. In a second step, we trained and tested on gaussian restless bandits (Daw et al., 2006, Chakroun et al., 2020) to directly compare RNN and human performance.
Results: RNNs with and without computation noise solved the four-armed restless bandit problem equally well across a range of test volatilities. A direct comparison to human performance revealed an overall similar accuracy, with some human participants significantly outperforming the networks.
Discussion: We discuss effects of computation noise in RNNs during reinforcement learning in volatile environments. Further analyses examine signatures of random and directed exploration. Furthermore, we discuss neural implications of this result and potential future research.
P009 - Functional decoding of thalamic nuclei: a database driven characterization of brain function.
1Humboldt Universität zu Berlin, Berlin, Germany; 2Institute of Neuroscience and Medicine (INM-1, INM-3), Research Centre Jülich, Jülich, Germany; 3Institute of Clinical Neuroscience and Medical Psychology, Heinrich Heine University, Düsseldorf, Germany,
The thalamus is a heterogeneous structure that consists of several distinct nuclei with unique connectivity patterns. Consensus views the thalamus as a relay station for brain-wide communication and studies have implicated the thalamus in sensorimotor as well as various cognitive functions. However, a precise characterization of thalamic nuclei from a cognitive neuroscience point of view is still missing.
We analyzed task-evoked data from a large amount of imaging studies as provided by the complementary databases BrainMap and Neurosynth for a functional decoding of thalamic function. The decoding was based on a novel thalamic parcellation which we obtained from a task-constraint meta-analytic connectivity based parcellation (CBP) in the BrainMap database.
The CBP resulted in a four and three clustering solution for the left and right hemisphere respectively, that showed reasonable accordance with cytoarchitectonic and anatomical maps of the thalamus. The post-hoc decoding through Bayesian reverse inference modeling, however, did reveal only a highly limited degree of specificity for cognitive terms across thalamic regions.
The surprising lack of specificity leaves the thalamus as a ‘cognitive blackbox’ for now. Next to methodological reasons such as the limited spatial resolution of functional MRI studies that blur boundaries of thalamic nuclei and reporting biases in the literature that omit thalamic activations, an isolated view on thalamic function might be disadvantageous: As a putative network hub, the thalamus might reveal its functional profile only in conjunction with interconnected brain areas. We will therefore discuss strategies for a systems-level decoding towards a network account of thalamic function.
P010 - High-quality ERPs in response to everyday sounds - captured with a smartphone
1Carl von Ossietzky Universität Oldenburg, Oldenburg, Germany · Department of Psychology; 2Jade University of Applied Sciences, Oldenburg, Germany · Institute for Hearing Technology and Audiology Master of Science
Mobile electroencephalography (EEG) technology allows studying brain processes in everyday life. One challenge in this emerging field is to relate brain activity to events in the environment. Here we describe how we use smartphone apps (developed by us) to record audio and EEG simultaneously with high temporal precision. The audio information is decomposed into features so that the recording complies with the legal requirements of recording audio in public. We show how to use this approach to study event-related potentials in relation to everyday sounds. We relate audio information in the from of spectral, amplitude, and onsets features to brain activity.
For validation, one participant was equipped with a mobile EEG amplifier and EEG cap, and microphones at the ears. EEG and audio were recorded concurrently on a smartphone. The participant listened to sounds played on a piano, to a story read by another person, and to a complex soundscape of a coffee shop. The sounds were recorded by the smartphone, turned into event codes, and saved together with the EEG. Based on these event codes, event-related potentials (EPR) were computed.
Results & Discussion
Formal timing tests show that EEG and audio are synchronized. For all conditions, clear auditory evoked potentials (P1, N1, P2) could be computed based on the auditory events. Using spectrum information, frequent and rare tones could be separated. ERPs in response to rare tones show a clear P3 component. Our approach demonstrates the feasibility of a pocketable lab: Real-world sensory-processing ERPs captured with smartphones.
P011 - Linear mixed models are superior to the arithmetic mean for stimulus norming
1Department of Clinical Psychology, Central Institute of Mental Health, Medical Faculty Mannheim, University of Heidelberg, Mannheim, Germany; 2Department of Addiction Behavior and Addiction Medicine, Central Institute of Mental Health, Medical Faculty Mannheim, University of Heidelberg, Mannheim, Germany; 3Department of Psychiatry and Psychotherapy, Central Institute of Mental Health, Medical Faculty Mannheim, University of Heidelberg, Mannheim, Germany; 4Heinrich-Heine-Universität, Düsseldorf
Introduction. Stimuli are one of the most important tools for psychologists. The more you know about the stimuli you are presenting in your experiment, the better your control over the experimental manipulation and the better your predictions about the dependent variable. This is especially crucial for biological psychology, where experiments are often resource-intensive and tight control over non-experimental variance is desired. Information about the properties of a given stimulus pool are commonly collected in norming studies (e.g. participants rate the valence of pictures). Usually, the stimulus properties of interest are extracted as the arithmetic mean across all participants. However, this estimate is biased in the scenario where individual participants are presented with only part of the stimulus pool (e.g. because the stimulus pool is very large). Linear mixed models explicitly take into account the variance that is introduced by the participants. Methods. We simulate stimulus norming data with varying degrees of participant, stimulus and error variance. Stimulus properties estimated by linear mixed models and the arithmetic mean are then compared to the true simulated stimulus properties. Results. Linear mixed models are superior in extracting the true stimulus properties, especially in the case where not every participant is presented with the whole stimulus pool. Discussion. Linear mixed models provide a more accurate rating of stimulus properties and should be preferred over the arithmetic mean.
P012 - Low-intensity transcranial focused ultrasound targeting the right prefrontal cortex leads to electroencephalographic midfrontal theta decreases which significantly predict approach behavior in a virtual T-maze task
1University of Würzburg, Würzburg, Germany; 2University of Arizona, Tucson (AZ), USA
Introduction: Low-intensity transcranial focused ultrasound (LITFUS) is a relatively new form of non-invasive neuromodulation with high precision for target selection and energy dosage, while avoiding side effects like headaches or skin irritations. Recent reviews highlighted its potential for basic research as well as clinical applications (Beisteiner & Lozano, 2020; Blackmore et al., 2019; Landhuis, 2017).
Methods: This double-blind within-subjects study (N = 155) utilized LITFUS targeting the right prefrontal cortex, which was found to positively enhance self-reported global mood (Hameroff et al., 2013; Reznik et al., 2020; Sanguinetti et al., 2020). The present study expanded these findings by using more specific self-report, and by adding a virtual T-maze task to measure approach behavior while recording electroencephalographic midfrontal theta (MFT), which has been associated with conflict experiences and behavior (Cohen & Donner, 2013). We hypothesized LITFUS would positively enhance self-reported mood, increase approach behavior and decrease MFT.
Results: Although no specific self-report changes were found, LITFUS led to significant MFT decreases, which significantly predicted increases in approach behavior.
Discussion: The LITFUS-induced MFT decreases and approach behavior increases confirmed our hypotheses. The absence of self-report effects might be due to our study’s focus on a task rather than self-reflection. This study expands the evidence for the impact of LITFUS on behavior and physiology, suggesting the promise of further basic and applied research, such as emotional and motivational disorders.
P013 - The gamma model analysis
Universität zu Köln
Research using the event-related potential (ERP) method to investigate cognitive processes has usually focused on the analysis of either individual peaks or the area under the curve as components of interest. These approaches, however, cannot analyse the substantial variation in size and shape across individual waveforms. The gamma model analysis (GMA) addresses these specific restrictions of the usually applied methods and enables the analysis of additional time-dependent and shape-related information on ERP components by fitting mathematically computed gamma probability density function (PDF) waveforms to an ERP.
The advantage of the GMA is demonstrated in a simulation study and a force production task. The different gamma model parameters were sensitive to various experimental manipulations across the empirical studies. Moreover, the GMA revealed several additional interrelated but non-redundant parameters compared to the classical methods, which were predictive of different aspects of behaviour, allowing for a more nuanced analysis of the cognitive processes. The GMA provides an elegant method for extracting easily interpretable indices for the rise and decline of the components that complement the classical parameters. This approach, therefore, provides a novel toolset to better understand the exact relationship between ERP components, behaviour, and cognition.
P014 - Training Factors for Tactile P300 Brain-Computer Interfaces
Universität Würzburg, Deutschland
INTRODUCTION: Brain-Computer Interfaces (BCIs) enable their users to interact with the environment based on brain activity, without relying on intact muscular function. Potential end-users thus include severely paralyzed patients.
Many BCIs are based on visually evoked P300 event-related potentials, leading to usability issues when eyesight or gaze control are impaired. Because of this limitation, vision independent alternatives have recently been developed. Specifically, a tactile paradigm originally intended for wheelchair control has been shown to be feasible and trainable for healthy users. Training is vital to ensure successful translation to potential end-users. This study aimed to confirm the trainability and explore which factors contribute to training effects.
METHODS: We analyzed performance and EEG data from 21 healthy participants across five tactile BCI sessions. Two experiments were included to identify potential training factors: Somatosensory sensitivity was assessed with a tactile discrimination task. A dual task condition explored whether training improved the BCI’s robustness against workload increase. Subjective workload was assessed via questionnaire.
RESULTS: We found a highly significant training effect on BCI accuracies (M = 78.7 % to 91.2 %) and on P300 amplitudes. No conclusive evidence for a role of workload was observed in the dual task condition, but the somatosensory sensitivity increased highly significantly between the first and last sessions. Notably, participants were able to discriminate between much smaller stimulus intensities after training.
DISCUSSION: The present study confirmed the trainability of the tactile BCI and provided first evidence of the importance of somatosensory sensitivity for training success.