Description:

Artificial intelligence methods show great promise in increasing the quality and speed of work with large astronomical and high-energy physics datasets. The high complexity of these methods leads to the extraction of dataset-specific, non-robust features, hence models do not generalize well across multiple datasets. As proof of concept, deep learning models are often trained on simulations with the prospect of being deployed and used on real data in the future. Unfortunately, this often leads to a substantial decrease in model performance on the real data. In this talk I will introduce "the domain shift problem" and why it appears in the sciences. Finally, I will introduce methods to overcome this problem and show several example studies performed by our group (ranging from galaxy morphology and strong gravitational lensing, to cosmology with galaxy distributions). With further development, these techniques will allow scientists to construct deep learning models that can successfully combine the knowledge from simulations and real data originating from multiple instruments. 

Alex Ciprijanovic, Wilson Fellow Associate Scientist, Fermilab

Host: Emma Alexander