Build Big meets Build Smart to Explore the Universe | Carolina Cuesta-Lazaro скачать в хорошем качестве

Build Big meets Build Smart to Explore the Universe | Carolina Cuesta-Lazaro

Speaker: Carolina Cuesta-Lazaro, Postdoctoral Fellow at the NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) at MIT Modern cosmology exemplifies the synergy between two complementary approaches in machine learning: scaling up with large models and datasets ("build big") versus incorporating targeted inductive biases for specific problems ("build smart"). Rather than choosing between these strategies, the most promising advances emerge from combining both. This talk presents three complementary projects that demonstrate this principle in action. First, I show how foundation models for science benefit from incorporating both simulated and observed data. By learning shared representations across these domains through alignment losses, we achieve robust simulation-based inference that remains reliable even under model misspecification. Second, I demonstrate scale-dependent anomaly detection using machine learning with cosmological inductive biases. By incorporating physical knowledge about scale dependence into the model’s architectures, we can detect deviations from standard models across different cosmological scales non-parametrically. This approach leverages both large observational datasets and physically-motivated architectural choices to identify potential new physics. Third, I explore using large language models for automated hypothesis generation in cosmology. Through a systematic evaluation framework, I show that LLMs can autonomously propose novel dark energy theories and implement them in existing physics codes like CLASS. While the approach shows promise, it also reveals current limitations, including implementation challenges for complex models and the tendency to improve fits through additional parameters rather than fundamental insights. Each project illustrates how the future of scientific discovery lies not in choosing between computational scale and inductive biases, but in thoughtfully combining both. This talk is part of the Liverpool Virtual Seminar Series on Data Intensive Science; more information can be found at https://indico.ph.liv.ac.uk/e/data_sc... #datascience #data #bigdata #education #cosmology