BayesFlow: Simulation Intelligence with Deep Learning

• Applicant: Dr. Paul-Christian Bürkner, since 12/2023
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 528702768

Simulation intelligence (SI) subsumes an emerging generation of scientific methods which utilize digital simulations for emulating and understanding complex real-world systems and phenomena. Recently, neural networks and deep learning have demonstrated a great potential for accelerating and scaling up SI to previously intractable problems and data sets. However, the availability of user-friendly software is still limited, which hampers the widespread and flexible use of modern SI methods. In this grant proposal, we focus on software for amortized Bayesian inference, which is an essential part of SI. The hallmark feature of amortized Bayesian inference is an upfront training phase (e.g., of a neural network), which is then amortized by a nearly instant fully Bayesian inference for an arbitrary number of data sets during test time. Concretely, we aim to advance the BayesFlow research software library into becoming the long-term, gold-standard software for amortized Bayesian inference. We define four specific objectives to achieve this goal: Firstly, we will increase the software's impact and usability by improving its internal structure, consistency, and interface, as well as enhancing its outreach metadata and documentation, such as API reference, quickstarts, and tutorials. Secondly, we will ensure sustainable quality by building efficient bug fixing routines, building test suites with maximal code coverage that are aligned with the inherent stochasticity of probabilistic models, and creating a comprehensive set of practically relevant benchmark models with a contribution guide for extensions. Thirdly, we will create conditions for further development by building a strong community around BayesFlow and amortized Bayesian inference more generally. This community will focus on both research and software aspects and help developers and users alike. Additionally, the proposal seeks to allow effortless switching between different deep learning backends (e.g., TensorFlow, PyTorch, JAX) to increase the scope of potential contributors and users. Finally, we will help refine the profile of a research software engineer by enabling them to collaborate with an interdisciplinary team of researchers who are working on various domain problems, such as computational biology, physics, statistics, and psychology. They will have the opportunity to delve deeper into these domain problems and contribute to solving substantial research questions through methodological and software innovations. Ultimately, by achieving these objectives, we will establish BayesFlow as a widely used and trustworthy software in the field of amortized Bayesian inference.

• DFG Programme: Research Grants
• Ehemaliger Antragsteller: Dr. Stefan Radev, until 12/2023