B03 – Parameter inference and model comparison in dynamical cognitive models
Objectives
In this project, we study dynamical models of eye-movement control—a representative example for cognitive modelling, where assumptions on information processing of the mind and brain are investigated to explain human behaviour. Since the human visual system is foveated (i.e., with high-resolution vision limited to a tiny region around the current gaze position), several gaze shifts per second (scan paths) are needed for visual information processing. We established the first examples for data assimilation of scan path models in scene viewing and reading. The next challenges are related to a broader range of model comparisons, to hierarchical modelling of interindividual differences, and to efficient algorithms in Approximate Bayesian Computation (ABC) for eye-movement models. Additionally, we will develop and host a model comparison platform for the field of eye movements in reading. With these goals, we expect to contribute innovative modelling work to the research field.
Data assimilation and model comparison are becoming increasingly important in cognitive modelling, since an increasing number of dynamical cognitive models are currently developed. We started with Bayesian parameter inference of the SceneWalk model for eye movements in scene viewing. In this model, the likelihood function could be computed, however, parallel computation on the level of individual fixation sequences needed to be implemented. In addition, we developed a new model, which implements the assumption of internal local and global attentional states and where Bayesian inference is based on data augmentation methods that allowed the use of conjugate priors and the construction of an efficient Gibbs sampler. In eye-movement control during reading, we investigated the SWIFT model, a model with stochastic internal states which implements a more complicated interaction between cognitive processes and oculomotor control (compared to the scene-viewing models above). SWIFT’s likelihood function could be decomposed into a spatial, pseudo-marginal likelihood and a temporal likelihood obtained by probability density approximation.
Based on the successful developments of likelihood-based methods for Bayesian inference for models in two domains (i.e., scene viewing and reading) and of new and/or improved models, we aim at model comparisons and developing hierarchical models in the second funding period. Model comparisons will require ABC and related methods, since we expect that likelihoods will not be available for many of the dynamical cognitive models in the research fields considered here. Problem-specific algorithms such as augmented Gibbs samplers are limited to models for which a data likelihood is explicitly available and of specific structure. Therefore, we will develop and implement more efficient and generically applicable likelihood and likelihood-free sampling methods which obey the principle of affine invariance. We will in particular rely on the preliminary work from Projects A06 and B04, respectively, on interacting particle samplers in order to develop affine invariant methods for ABC and statistical inference for generative cognitive models with maximum mean discrepancy (MMD).
In addition to our own models, we will work with models such as the E-Z Reader model for eye guidance in reading and the LATEST model for scene viewing. As a consequence, there are challenging problems for objective model comparisons, which clearly require advanced inference methods such as ABC and MMD in combination with domain-specific insight.
The project will closely interact with projects A06 on Bayesian inference and project B09 on cognitive modelling.
Preprints
Chen, Y., Huang, D.Z., Huang, J., Reich, S., & Stuart, A.M. (2024). Efficient, Multimodal, and Derivative-Free Bayesian Inference With Fisher-Rao Gradient Flows. arXiv:2406.17263
Engbert, R. and Rabe, M. M. (2023). Tutorial on dynamical modeling of eye movements in reading. doi: 10.31234/osf.io/dsvmt
Irwin, B. and Reich, S. (2023). EnKSGD: A class of preconditioned black box optimization and inversion algorithms. arXiv:2303.16494.
Schwetlick, L. and Reich S. and Engbert R. (2023). Bayesian Dynamical Modeling of Fixational Eye Movements. arXiv:2303.11941.
Rabe, M. M., Paape, D., Mertzen, D., Vasishth, S., and Engbert, R. (2023). SEAM: An integrated activation-coupled model of sentence processing and eye movements in reading. arXiv:2303.05221
Huang, D.Z., Huang, J., Reich, S., and Stuart, A.M. (2022). Efficient derivative-free Bayesian inference for large-scale inverse problems. arXiv:2204.04386.
Seelig, S., Risse, S., and Engbert, R. (2020). Predictive modeling of the influence of parafoveal informationprocessing on eye guidance in reading. doi:10.31234/osf.io/vbmqn
Publications
Irwin, B., & Reich, S. (2024). EnKSGD: A class of preconditioned black box optimization and inversion algorithms. SIAM Journal on Scientific Computing, 46, A2101-A2122. doi: 10.1137/23M1561142
Yadav, H., Smith, G., Reich, S., and Vasishth, S. (2023). Number feature distortion modulates cue-based retrieval in reading. Journal of Memory and Language, Vol. 129, 104400. doi: 10.1016/j.jml.2022.104400
Schwetlick, L.; Backhaus, D. & Engbert, R. (2022). A dynamical scan-path model for task-dependence during scene viewing. Psychological Review, American Psychological Association (APA). doi: 10.1037/rev0000379
Huang, D.Z., Huang, J., Reich, S., and Stuart, A.M. (2023). Efficient derivative-free Bayesian inference for large-scale inverse problems. Inverse Probelms, Vol. 38, 125006. doi: 10.1088/1361-6420/ac99fa, arXiv:2204.04386
Engbert, R., Rabe, M. M., Schwetlick, L., Seelig, S. A., Reich, S., Vasishth, S. (2022). Data assimilation in dynamical cognitive science. Trends in Cognitive Sciences, 26(2), 99-102. doi:10.1016/j.tics.2021.11.006.
Malem-Shinitski, N., Ojeda, C., and Opper, M. (2022). Variational Bayesian Inference for Nonlinear Hawkes Process with Gaussian Process Self-Effects. Entropy, 24(3), 356. doi: 10.3390/e24030356.
Rabe, M. M., Chandra, J., Krügel, A., Seelig, S. A., Vasishth, S., & Engbert, R. (2021). A Bayesian approach to dynamical modeling of eye-movement control in reading of normal, mirrored, and scrambled texts. Psychological Review doi:10.1037/rev0000268, psyarXiv
Schwetlick, L., Rothkegel, L.O.M., Trukenbrod, H.A., Engbert, R. (2020). Modeling the effects of perisaccadic attention on gaze statistics during scene viewing. Communications Biology, 3, 727. doi: 10.1038/s42003-020-01429-8
Malem-Shinitski, N., Opper, M., Reich, S., Schwetlick, S., Seelig S. A., & Engbert, R (2020). A Mathematical Model of Exploration and Exploitation in Natural Scene Viewing. PLoS Computational Biology. doi:10.1371/journal.pcbi.1007880
Seelig, S. A., Rabe, M. M., Malem-Shinitski, N., Risse, S., Reich, S., and Engbert, R. (2020). Bayesian parameter estimation for the SWIFT model of eye-movement control during reading. Journal of Mathematical Psychology, 95, 102313. doi:10.1016/j.jmp.2019.102313; arXiv: 1901.11110.
Engbert, R., Rabe, M.M., Seelig, S.A., & Reich, S. (2019). Bayesian parameter estimation for dynamical models of eye-movement control using adaptive Markov Chain Monte Carlo simulations. Forschung im HLRN-Verbund 2019.
Malem-Shinitski, N., Seelig, S. A., Reich, S. and Engbert, R. (2019). Bayesian inference for an exploration-exploitation model of human gaze control. Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany (extended abstract). doi:10.32470/CCN.2019.1246-0
Seelig, S. A., Rabe, M. M., Malem-Shinitski, N., Reich, S., Engbert, R. (2019). Parameter estimation for the SWIFT model of eye-movement control during reading. Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany (extended abstract) doi:10.32470/CCN.2019.1369-0
Schütt, H. H., Rothkegel, L. O. M., Trukenbrod, H. A., Reich, S., Wichmann, F. A. and Engbert, R. (2017). Likelihood-based parameter estimation and comparison of dynamical cognitive models. Psychological Review, 124, 505-524. doi:10.1037/rev0000068