New research shows that transfer learning can reduce the number of expensive cosmological simulations needed to train a neural network by more than tenfold. However, prior knowledge of the accepted cosmological model can also bias the system, making it difficult for it to detect phenomena that deviate from it.
Intelligence can accelerate the pursuit of new physics, but sometimes it knows too much to see what is right under its nose.
Artificial intelligence could make it possible to search for new laws of physics in less time and money, according to a new study published in JCAP. But the study also points to a downside. In some situations, AI can become so dependent on its previous training that it has trouble identifying truly new phenomena.
AI has become an important tool in cosmology, helping researchers analyze vast amounts of data about the universe. But exploring ideas that go beyond the Standard Cosmological Model (ΛCDM) is still a very expensive computational challenge.
ΛCDM successfully explains many observable properties of the universe, including its expansion and the large-scale dispersion of galaxies, but scientists believe it does not tell the whole story. Recent observations suggest that phenomena such as massive neutrinos, altered gravity, and evolving dark energy could reveal physics beyond the current model.
To explore these possibilities, researchers need to create a huge number of detailed simulations of virtual universes, each based on different physical assumptions. Creating these simulations often requires a lot of computing power and time.
Transfer learning offers a faster route
The researchers tested whether a machine learning approach called learning transfer could reduce the burden.
Transfer learning allows an AI system to apply knowledge gained from one task to help it learn another task more effectively. Instead of starting from scratch, the AI builds on what it has already learned.
In this study, the team first trained a neural network using simulations based on ΛCDM. This initial training process gave the AI a foundation before exposing it to more complex cosmological models that include possible new physics.
"It's basically a shortcut," explains cosmologist Adrian Bayer, one of the study's authors. "Usually people train the AI directly on the most computationally expensive simulations. Instead, we use simpler, less expensive ΛCDM simulations to give the AI an idea of what's going on, and only then move on to the more complex models."
Bayer compares the process to learning from books. "First you read a basic book to get an idea of the knowledge, then you move on to a really complicated book."
According to lead author Veena Krishnaraj, with this method, artificial intelligence doesn't have to "digest the whole thing at once."
The strategy has proven to be very effective. In some cases, the transfer of learning has reduced the number of expensive simulations needed by more than tenfold.
When prior knowledge becomes a problem
The study also revealed a less obvious problem called negative transference.
In Bayer's analogy to books, suppose a medical student studies introductory materials and then encounters a rare disease that resembles a common one. Existing knowledge is usually helpful, but sometimes it can lead to the wrong conclusion.
A similar problem can arise in artificial intelligence systems. Some signals produced by new physics can look very similar to patterns that the AI has already learned from the Standard Cosmological Model. When this happens, the AI may interpret the new information through the lens of its previous training, making it harder to detect anything that is truly different.
The researchers saw this phenomenon when they studied simulations involving massive neutrinos. Some of the observed results of the neutrino mass closely resemble changes associated with an existing parameter in the ΛCDM model called σ8, which measures the strength of the clumping of matter throughout the universe.
Because the two phenomena can look so similar, the pre-trained neural network initially had difficulty distinguishing between them.
The article was published inJournal of Cosmology and Astroparticle Physics (JCAP), June 2026. (arXiv)
More of the topic in Hayadan:
One response
Over Fitting – Overfitting is not a new issue in machine learning.
It's a trade-off. The more points you rely on, the better the model gets at locating the points and the **worse** at making predictions based on information it hasn't seen.
https://en.wikipedia.org/wiki/Overfitting
https://he.wikipedia.org/wiki/%D7%94%D7%AA%D7%90%D7%9E%D7%AA_%D7%99%D7%AA%D7%A8