To discover new physics, AI may need to 'unlearn' the old one

Key Points:

• A new study in the Journal of Cosmology and Astroparticle Physics shows that transfer learning, a machine-learning technique where AI reuses knowledge from simpler tasks, can significantly reduce the computational cost of testing new physics beyond the standard cosmological model (ΛCDM).
• Researchers pretrained neural networks on less expensive ΛCDM simulations before adapting them to more complex models, reducing the number of costly simulations needed by over a factor of 10 in some cases.
• However, the study also identified a risk called negative transfer, where pretrained AI can misinterpret new physics effects that resemble known patterns, as seen with massive neutrinos mimicking existing ΛCDM parameters, complicating accurate recognition of genuinely new phenomena.
• This phenomenon is driven by physical degeneracies where different parameters produce similar observable effects, highlighting the need for careful mitigation strategies when using transfer learning in cosmology.
• The approach has so far been tested on simulations and holds promise for analyzing upcoming high-precision cosmological survey data, though researchers caution that pretraining may both accelerate inference and potentially hinder the discovery of new physics.