Research

The central research goal of the group is to advance our understanding of the neurocomputational foundations of human behavior through the continuous development and experimental evaluation of probabilistic data-analysis models. In the long term, we aim to vertically link the computational, algorithmic, and implementational levels in Marr's framework of cognitive processes. Across the different research projects, we pursue an interdisciplinary approach that combines mathematical modeling, computer simulation, and empirical evaluation based on human behavioral and neuroimaging data, primarily EEG.

Agent-Based Behavioral Modeling

In a series of studies, we investigate the neuroalgorithmic foundations of sequential human decision-making under uncertainty. The central methodological paradigm combines controlled behavioral experiments with agent-based behavioral modeling, based on the statistical embedding of concepts from agent-based artificial intelligence, partially observable Markov decision processes, and modern reinforcement learning. Medium-term goals include, first, the formulation and implementation of a general statistical framework that allows diverse agent models to be evaluated in a standardized way using human behavioral experimental data. Second, highly controlled experimental adaptations of computer games are used to achieve higher empirical validity.

Selected Publications

Usee F, Schmidt S, Melzig CA, Ostwald D (2025) Agent-based behavioral modeling of human associative learning in a complex approach-avoidance conflict task  PsyArXiv  Data & Code  Computational Brain and Behavior

Horvath L, Colcombe S, Milham M, Ray S, Schwartenbeck P, Ostwald D (2021) Human belief state-based exploration and exploitation in an information-selective reversal bandit task  BioRxiv  Data & Code  Computational Brain & Behavior

Probabilistic Analysis Methods for Neuroimaging Data

In a second research focus, the group validates and develops new probabilistic methods for analyzing neuroimaging data. The focus is, on the one hand, on the theoretical properties and neuroscientific potential of variational inference. The primary goal in this area is the theoretical and simulation-based validation of the Variational-Laplace methodology that is widely used in the neuroimaging community. On the other hand, the group aims to further develop random-field-theory-based activity inference. For example, power and sample-size analyses under both frequentist and Bayesian inference are to be developed and validated using computationally intensive simulations. In addition to their data-analytic potential far beyond applications in cognitive neuroscience, both research foci also address current discussions about the validity and reproducibility of scientific findings.

Selected Publications

Ostwald D, Schneider S, Bruckner R, Horvath L (2021) Random field theory-based p-values: A review of the SPM implementation  arXiv  Data & Code

Neurocognition

The goal of the neurocognition research project is to move beyond purely correlational analyses in opening up Marr's implementational level and to arrive at a biophysical theory of neuropsychological processes and their EEG-based evaluation. Initial foundations for this research project have been laid in recent years. Medium-term goals include, first, the computational decomposition of neural correlates of human decision processes in visually highly reduced decision tasks, such as bandit tasks, through a series of computational EEG studies. Second, by further developing experimental and data-analytic approaches, we aim to better understand neural EEG correlates of decision processes in visually more demanding scenarios with higher external validity, such as computer games.

Selected Publications

Ostwald D, Usee F (2021) An induction proof of the backpropagation algorithm in matrix notation  arXiv  Code

Ostwald D, Starke L (2016) Probabilistic delay differential equation modelling of event-related potentials  PDF  Code  NeuroImage