Helmholtz AI researchers and collaborators introduce a gradient-based causal approach for the problem of Bayesian optimal experimental design even in batch settings.
Unlocking Experimental Design: Broadening Causal Discovery Applications
In simple terms, experimental design is about conducting the next experiments to identify the underlying system. In many applications, scientists come up with hypotheses and design experiments to test these hypotheses, but in many applications we need to derive these hypothesis from data and one wants to identify the system with as few experiments as possible due to costs and risks involved. The overall process can be very complex and computationally challenging.
Researchers have used a Bayesian framework to optimally explore the space of possible experiments, and in a data-driven way proposes the experiments to run next. This is especially important for applications with huge search spaces and limited human derived hypothesis e.g. for discovering underlying gene networks or discovering novel materials.
In this work in a collaboration between Oxford University, Microsoft Research and Helmholtz Munich, Panagiotis Tigas, Yashas Annadani, Desi R. Ivanova, Andrew Jesson, Yarin Gal, Adam Foster and Stefan Bauer, propose a new method that uses causality to make the process more sample efficient. They introduce a gradient-based causal approach for the problem of Bayesian optimal experimental design in a batch setting without relying on specific assumptions about the causal model.
Overall, their approach opens up new possibilities for experimental design in various applications compared to previous methods, making it easier to discover causal relationships in complex systems and applying them e.g. to large-scale CRISPR experiments as routinely explored in many experiments at Helmholtz Munich and around the globe.
This work is a collaboration between University of Oxford, KTH Stockholm, Helmholtz AI, Microsoft Research and TU Munich, and has been published in Proceedings of the 40 the International Conference on Machine Learning in Honolulu, Hawaii, USA.
Original publication
Tigas, P., Annadani, Y., Ivanova, D.R., Jesson, A., Gal, Y., Foster, A., Bauer, S. (2023). Differentiable Multi-Target Causal Bayesian Experimental Design. Proceedings of the 40th International Conference on Machine Learning. https://proceedings.mlr.press/v202/tigas23a.html