Simulation-based inference for complex models
Most models for real-life applications, e.g., stochastic models for systems biology, finance and population dynamics, have an intractable likelihood-function, that is, the likelihood is not available in closed form, thus preventing standard inference approaches to be carried out. Dr Umberto Picchini develops algorithmic strategies for inferring model parameters when customary inference methods are not applicable.
Picchini is particularly interested in methods known as “simulation-based inference” (SBI) or “likelihood free inference” methods, where the ability to simulate from computer models is the only requirement. However, these can be computationally demanding and he proposed ways to make these SBI methods more efficient.
His SBI methods have been used for inference in, e.g., Markov jump processes for systems biology [1], cell-movements models [2], stochastic differential equations for neural activity [3], and tumor growth [4], where likelihood-free versions of Markov chain Monte Carlo, sequential Monte Carlo, and neural-networks-based [5-6] approaches have been exploited.
References
[1] S. Persson, N. Welkenhuysen, S. Shashkova, S. Wiqvist, P. Reith, G. W. Schmidt, U. Picchini, M. Cvijovic (2022). Scalable and flexible inference framework for stochastic dynamic single-cell models, PLOS Computational Biology, 18(5):e1010082.
[2] U. Picchini and M. Tamborrino (2022). Guided sequential ABC schemes for intractable Bayesian models, arXiv:2206.12235.
[3] Wiqvist, S., Golightly, A., McLean, A. T., & Picchini, U. (2021). Efficient inference for stochastic differential equation mixed-effects models using correlated particle pseudo-marginal algorithms. Computational Statistics & Data Analysis, 157, 107151.
[4] U. Picchini and J. Forman (2019). Bayesian inference for stochastic differential equation mixed effects models of a tumor xenography study, Journal of the Royal Statistical Society (Series C), 68(4), 887-913,
[5] S. Radev, M. Schmitt, V. Pratz, U. Picchini, U. Köthe, P. Bürkner (2023). JANA: jointly amortized neural approximation of complex Bayesian models. The 39th Conference on Uncertainty in Artificial Intelligence (UAI 2023), vol 216, pp. 1695-1706.
[6] S. Wiqvist, P-A. Mattei, U. Picchini and J. Frellsen (2019). Partially Exchangeable Networks and architectures for learning summary statistics in Approximate Bayesian Computation. Proceedings of the 36th International Conference on Machine Learning, PMLR 97:6798–6807.