New research indicates that transfer learning can significantly accelerate the search for new physics, reducing the need for expensive simulations. However, the reliance on established patterns may cause AI to overlook genuinely novel phenomena, according to a study published in the Journal of Cosmology and Astroparticle Physics (JCAP).
The standard model of cosmology, known as ΛCDM, explains many universe features but is not comprehensive. New observations have raised questions regarding concepts such as massive neutrinos, modified gravity, and evolving dark energy. Investigating these requires extensive computer simulations, which are both computationally expensive and time-consuming.
The research team aimed to determine if transfer learning could enhance simulation efficiency. Transfer learning allows an AI to apply knowledge from simpler tasks to more complex ones, thus lowering costs. Initially, the AI was trained on basic ΛCDM simulations before transitioning to more complex models that incorporated potential new physics.
Adrian Bayer, a co-author from the Flatiron Institute and Princeton University, described this method as a shortcut to traditional AI training. “Usually people train the AI directly on the most computationally expensive simulations. What we do instead is first use simpler and less expensive ΛCDM simulations to give the AI an idea of what’s happening,” Bayer said.
This pretraining strategy helped the AI manage complexity without being overwhelmed. The study found that transfer learning reduced the number of costly simulations needed by over a factor of ten in some cases.
The researchers also identified a challenge known as negative transfer, which occurs when AI misinterprets new information based on its pre-existing knowledge. The AI often struggled to discern new effects when they resembled patterns aligned with existing ΛCDM parameters. This was evident in simulations involving massive neutrinos, where the AI faced difficulty differentiating new signatures from those it had already associated with known parameters.
Veena Krishnaraj, the study’s lead author, explained that the negative transfer results from underlying physical degeneracies in the models. “Different physical processes can produce very similar observable signatures,” she noted, indicating the need for caution in interpreting AI findings.
The study emphasizes both the potential benefits and limitations of transfer learning in the physics domain. While pretraining can expedite data analysis, it may hinder the AI’s capability to recognize groundbreaking discoveries. The next phase will involve applying the transfer learning technique to real astronomical observations.
The researchers anticipate that transfer learning will be pivotal for upcoming cosmological surveys set to gather high-precision data. The paper titled “Transfer Learning Beyond the Standard Model,” authored by Veena Krishnaraj and colleagues, is now published in JSTAT.
