Efficient Simulation-Based Inference from Noisy Scientific Data

A key problem in data-driven science is fitting simulation model parameters to experimental data.

Solving this simulation-based inference problem efficiently and reliably is critical towards realising novel scientific discoveries. We consider the setting where the simulation model is stochastic, and experimental data noisy. We also assume that the likelihood function of the simulator is either unknown or intractable.

The current state-of-the-art herein is dominated by deep learning, where a mapping between observed data and corresponding simulator parameters is learned. The project will address two major challenges.

First, existing neural network approaches either do not consider the underlying aleoteric and epistemic uncertainties, or do not scale well with increasing problem complexity.

Here we propose to build novel Bayesian architectures based on variational inference that are uncertainty-aware, and will scale to infer hundreds of parameters.

Second, the networks require substantial training data, which is simulated from a prior distribution over the parameters.

Here we propose to design a data-efficient active learning approach combining space-filling exploration of the parameter space, with exploitation of inference-critical areas. The project will fund 2 postdoctoral positions for 2-years each, potentially bringing new talent to Sweden.

The scientific applications come from 3 separate existing collaborations in studying gene regulation, particle physics and photonics.