Researchers find AI can miss cosmology signals from negative transfer
A new study shows that AI could really shake up how we explore the universe, especially in cosmology.
Researchers trained an AI on the standard model of the cosmos (LCDM), and it handled basic data well.
But when things got tricky, like with dark energy or modified gravity, the AI stumbled because of something called "negative transfer," where it mixes up patterns and misses new discoveries.
Research team plans realistic galaxy tests
Co-author Adrian E. Bayer pointed out that figuring out when to trust these AI tools is key, since mistakes could lead scientists down the wrong path.
The team tried transfer learning to help the AI adapt without many new simulations, but this brought in biases that made spotting unique phenomena harder.
Next up: they're planning tests with more realistic galaxy data (including noise and uncertainties) to make sure future cosmic research gets a reliable boost from AI.