Watch our interview with Aleksandra on Youtube.
Aleksandra Ćiprijanović is a postdoctoral research associate at Fermi National Accelerator Laboratory, where she works for the Scientific Computing Division on the High-velocity Artificial Intelligence project. More broadly, she is interested in applying machine learning and data science to astronomy, cosmology, and high-energy physics.
Related
Show Notes
- Like many future scientists, Aleksandra grew up being fascinated by both dinosaurs and stars. Alas for the field of paleontology it was her interest in space that stuck, and she went on to study computer science, mathematics, and astronomy in her native Serbia.
- Her early interests in astrophysics included high-energy particle physics, focused on questions like how particles formed in the moments after the big bang.
- I asked Aleksandra if there were any major concepts she wish were imported out of astrophysics to the broader thinking public. She told me that her field studies the large-scale structures of the universe, and it really drives home how a finite set of forces, like gravity, shape phenomena on the Earth, in the heavens, and everywhere.
- I also wanted to know what the evidentiary basis is for our claims about the structure of the cosmos and the physical processes going on inside galaxies and stars.
- The answer isn't very surprising: it mostly comes down to building better instruments and following very, very careful chains of reasoning, inference, and proof.
- We then turned to discussing her application of machine learning to astrophysics. Much of it involves using complex deep learning models like convolutional neural networks to identify objects in images.
- At this stage most data gathering, cleaning, and tagging is highly manual, and that's an area where improvements could be made in the future. Aleksandra is working with colleagues to do just that at present.
- A key technique in this effort is called 'domain adaptation', which is also used in e.g. training autonomous vehicles.
- The key idea is to train the model on one dataset but only allow it to learn features which are also present in another dataset of interest. This forces it to only learn domain invariant features, allowing to better generalize.
- Today, Aleksandra is looking to use deep learning to understand galactic mergers, the process by which two galaxies collide and coalesce.
- Why does this matter? Because it can help us to better understand the origins of different galactic shapes, how galaxies evolve, and the distribution of stars and elements in the universe.
- The penultimate leg of the conversation was spent on dark matter, and I wanted to know why anyone believes that dark matter exists. She replied that given the velocities with which stars rotate around the galactic center, and given the observable masses of stars in galaxies, there just shouldn't be enough gravitational power to keep them together. But they clearly are, so physicists have postulated an additional form of matter which binds them together.
- Finally, we spoke about Aleksandra's involvement with the deep skies project, which aims to broaden the application of machine learning and deep learning to fields like astrophysics and cosmology.
Share this episode.
