Description
The management of neurodegenerative conditions such as Parkinson’s and Alzheimer’s disease are becoming a ever more prevalent challenge in modern society. However, current population-level models often fail to account for the significant heterogeneity in how individual patients present and respond to therapy [1, 2]. Amortized Bayesian Inference (ABI) [3, 4], is a promising paradigm that “front-loads” computational costs during a simulation-based training phase. Once a model is trained, it enables nearly instantaneous inference on new patient data (e.g., EEG or fMRI), potentially making personalized treatment selection feasible within routine clinical workflows.
Research Problem
In its current state ABI still has major bottlenecks to overcome. One particular one is the use of non-informative priors as this introduces two key challenges [3, 5]:
• Computational Inefficiency: Uniform or uninformative priors waste simulation capacity on parameter spaces that have no biological relevance.
• Simulator-Reality Mismatch: Standard deep neural density estimators risk converging on biologically implausible regimes, exacerbating systematic bias when models are misspecified.
This thesis investigates whether integrating domain-expert prior knowledge into deep neural density estimators can act as an effective regularizer. You will test the hypothesis that informed priors improve simulation efficiency during training while preventing the model from falling into unphysiological parameter regimes. Thus, ultimately building more robust, clinically viable inference pipelines.
Desired Background and Skills
Applicants should have strong analytical and programming skills, preferably in Python. A background in computational neuroscience, Bayesian statistics, signal processing, applied mathematics, deep learning, or bioinformatics is highly desirable. The ideal candidate thrives in an interdisciplinary environment and is motivated to engage with mathematical theory, neural data, and computational modeling.
To apply, please submit a short statement of motivation (addressing both the research group and topic), your CV, and your academic transcripts to computationalneurology@rub.de.
References
[1] G. B. Frisoni, N. C. Fox, C. R. Jack, P. Scheltens, and P. M. Thompson. “The clinical use of structural MRI in Alzheimer disease”. In: Nature reviews. Neurology 6.2 (Feb. 2010), pp. 67–77. issn: 1759-4758. doi: 10.1038/nrneurol.2009.215.
[2] K. E. Stephan, D. R. Bach, P. C. Fletcher, J. Flint, M. J. Frank, K. J. Friston, A. Heinz, Q. J. M. Huys, M. J. Owen, E. B. Binder, P. Dayan, E. C. Johnstone, A. Meyer-Lindenberg, P. R. Montague, U. Schnyder, X.-J. Wang, and M. Breakspear. “Charting the landscape of priority problems in psychiatry, part 1: classification
and diagnosis”. en. In: The Lancet Psychiatry 3.1 (Jan. 2016), pp. 77–83. issn: 22150366. doi: 10.1016/S2215-0366(15)00361-2.
[3] K. Cranmer, J. Brehmer, and G. Louppe. “The frontier of simulation-based inference”. In: Proceedings of the National Academy of Sciences 117.48 (Dec. 2020), pp. 30055–30062. doi: 10.1073/pnas.1912789117.
[4] S. T. Radev, U. K. Mertens, A. Voss, L. Ardizzone, and U. K¨othe. BayesFlow: Learning complex stochastic models with invertible neural networks. en. arXiv:2003.06281 [stat.ML]. Dec. 2020. doi: 10.48550/arXiv.
2003.06281.
[5] P. J. Gon¸calves, J.-M. Lueckmann, M. Deistler, M. Nonnenmacher, K. ¨ Ocal, G. Bassetto, C. Chintaluri, W. F. Podlaski, S. A. Haddad, T. P. Vogels, D. S. Greenberg, and J. H. Macke. “Training deep neural density estimators to identify mechanistic models of neural dynamics”. In: eLife 9 (Sept. 2020). Ed. by J. R. Huguenard,
T. O’Leary, and M. S. Goldman, e56261. issn: 2050-084X. doi: 10.7554/eLife.56261.