In a study published earlier this month in the Journal of Cosmology and Astroparticle Physics, cosmologists trained an AI neural network on simulations of the ΛCDM—the Standard Model of cosmology (hereafter the Standard Model). Then, the team tested whether this pre-training would help or harm the AI in subsequent investigations of other outstanding problems in cosmology and astrophysics. Although AI showed some promise, it developed biases that proved detrimental to the discovery of new physics.
“This study is a good example of how AI can help science advance faster when it is used in a structured way,” study co-author Adrian E. Baer, a cosmologist at the Flatiron Institute and Princeton University, told Gizmodo. “At the same time, the study is a reminder that acceleration and understanding must go together.”
expensive truth
Cosmological breakthroughs are expensive and time-consuming. As Will Percival, co-spokesman for the Dark Energy Spectroscopic Instrument (DESI), told Gizmodo in April, preparing the dataset for scientific analysis involves creating fake universes and galaxies and then running the simulations as sanity checks. These processes are important to draw any serious conclusions from advanced observations.
But simulations of models beyond the Standard Model — extensions that include massive neutrinos, evolved dark energy or modified gravity — are also very expensive, Bayer told Gizmodo. At the same time, testing these alternative scenarios, whether or not they are ultimately correct, is important for furthering our understanding of the universe. That was the practical motivation that led Bayer to seek “methods that can learn efficiently without requiring large new simulation suites for every scenario.”
Bumpy transfers?
For the experiment, the team used a machine learning strategy called transfer learning. In this approach, a model first learns from a task or dataset—simulations of the Standard Model—and applies this knowledge to learn related tasks or extended versions of the Standard Model that include promising ideas for new physics.
According to Bayer, AI performed quite well in understanding the standard model based on short, less expensive simulations. However, it began to struggle when the new physics “overlapped with the directions in which it had already learned [the standard model] Parameter space,” he said. This phenomenon, called negative transfer, emerged when AI became biased and was not able to distinguish between two different physical effects that produced similar patterns in the data. So instead of naturally discovering something new, the AI relied on what it had already learned, causing it to miss potential clues that hinted at physics beyond the Standard Model.
“The negative transfer result is fascinating because it shows that the model is not failing randomly,” Bayer said. “To reliably use AI in future cosmological analyses, it is very important to understand when transfer learning helps and when it reinforces those distortions.”
AI and cosmology
For Bayer, the latest findings confirm the new notion that AI can be helpful, but human experts must follow its calculations carefully to understand and pursue relevant questions.
“Transfer learning could give AI a powerful head start, allowing us to test many more ideas about the universe that otherwise wouldn’t be practical,” he said. “But if a model carries knowledge from one setting to another, we need to understand what it does – when that knowledge helps and when it can mislead.”
Next, Bayer and colleagues plan to conduct similar experiments in settings that “more closely resemble actual survey data” that include “galaxy formation uncertainties, survey masks, and noise.” Additionally, the team wants to explore which cosmological inquiries might benefit most from transfer learning.
<a href