Хиймэл оюун ухааныг сансар судлалын тооцоололд ашиглах нь судалгааны ажлыг хурдасгах боломжтой ч шинжлэх ухааны шинэ нээлт хийхэд сөрөг нөлөөтэй байж болохыг судалгаагаар тогтоожээ.
Cosmology and Astroparticle Physics сэтгүүлд нийтлэгдсэн судалгаагаар эрдэмтэд хиймэл оюун ухааныг сансар судлалын стандарт загвар болох ΛCDM-ийн симуляцаар сургаж, улмаар түүнийг сансар судлалын бусад тулгамдсан асуудлыг шийдвэрлэхэд ашиглаж үзжээ. Флатироны хүрээлэн болон Принстоны их сургуулийн судлаач Адриан Э. Байер нарын үзэж буйгаар, хиймэл оюун ухаан нь стандарт загварын хүрээнд сайн ажиллаж байсан ч шинэ физикийн үзэгдлийг илрүүлэхэд өрөөсгөл хандлагатай болж байв.
Сансар судлалын томоохон нээлтүүд нь маш их цаг хугацаа, зардал шаарддаг бөгөөд эрдэмтэд үүнийг хөнгөвчлөхийн тулд “transfer learning” буюу мэдлэг шилжүүлэх аргыг туршжээ. Энэ аргаар хиймэл оюун ухаан нь стандарт загвараас сурсан мэдлэгээ илүү өргөн хүрээний судалгаанд ашиглах ёстой байв. Гэвч хиймэл оюун ухаан нь өмнө нь сурсан мэдээлэлдээ хэт найдаж, ижил төстэй физик үзэгдлүүдийг ялгаж чадахгүй “сөрөг шилжилт” хэмээх алдаа гаргасан байна.
https://iopscience.iop.org/article/10.1088/1475-7516/2026/06/026
https://adrianbayer.github.io/
Адриан Э. Байерын тэмдэглэснээр, хиймэл оюун ухаан нь судалгааны ажлыг хурдасгах хүчирхэг хэрэгсэл боловч түүний гаргасан үр дүнг хүн-шинжээчид нарийн хянах шаардлагатай юм. Судлаачид цаашид хиймэл оюун ухааныг бодит судалгааны өгөгдөл, галактикийн үүсэлтэй холбоотой тодорхой бус байдал, чимээ шуугиантай нөхцөлд хэрхэн ажиллахыг туршихаар төлөвлөж байна.
Дэлгэрэнгүйг эх сурвалжаас харах
Эх сурвалжийг нээх ↓
When cosmology makes headlines, we often see fancy images of cosmic maps and supernovas. But in reality, scientists have to sift through hundreds or thousands of calculations and simulations for months or years. In an effort to reduce this burden, some scientists have turned to AI—but, as a new study finds, the pros and cons are quite nuanced.
In a study published earlier this month in the Journal of Cosmology and Astroparticle Physics, cosmologists trained an AI neural network on simulations of ΛCDM—the standard model of cosmology (hereafter the standard model). Then, the team tested whether this pre-training would help or hurt the AI’s subsequent investigations into other outstanding problems in cosmology and astrophysics. Although the AI did show some promise, it developed biases that ended up being detrimental to finding new physics.
The study is a “nice example of how AI can help science move faster when it is used in a structured way,” Adrian E. Bayer, the study’s co-author and a cosmologist at the Flatiron Institute and Princeton University, told Gizmodo. “At the same time, the study is a reminder that acceleration and understanding have to go together.”
The costly truth
Cosmological breakthroughs tend to be costly and time-consuming. As Dark Energy Spectroscopic Instrument (DESI) co-spokesperson Will Percival told Gizmodo back in April, preparing datasets for scientific analysis involves the creation of mock universes and galaxies and then running simulations as sanity checks. These processes are vital for drawing any serious conclusions from advanced observations.
But simulations of models beyond the standard model—extensions that involve massive neutrinos, evolving dark energy, or modified gravity—are also very expensive, Bayer told Gizmodo. At the same time, testing these alternative scenarios, regardless of whether they end up being right, is critical in advancing our understanding of the cosmos. That practical motivation was what led Bayer to look for “methods that can learn efficiently without requiring huge 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 one task or dataset—simulations of the standard model—and applies this knowledge to learn a related task or extended versions of the standard model that include promising ideas for new physics.
According to Bayer, the AI performed quite well in terms of understanding the standard model based on fewer, less costly simulations. However, it began to struggle when new physics “overlaps with directions it has already learned in [the standard model] parameter space,” he noted. This phenomenon, called negative transfer, emerged as the AI became biased and wasn’t able to distinguish between two different physical effects that produce similar patterns in the data. So instead of spotting something inherently new, the AI relied on stuff 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 added. “Understanding when transfer learning helps and when it reinforces those degeneracies is very important for using AI reliably in future cosmological analyses.”
AI and cosmology
For Bayer, the latest findings affirm the not-so-novel notion that AI can be helpful, but human experts must carefully follow its calculations to understand and pursue relevant questions.
“Transfer learning can give AI a powerful head start, allowing us to test many more ideas about the universe than would otherwise be practical,” he said. “But if a model carries knowledge from one setting into another, we need to understand what it has carried over—when that knowledge helps and when it might 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 could benefit most from transfer learning.
