Description
Machine learning is increasingly shaping cosmological analysis, though
its impact depends critically on careful validation and physically interpretable use.
In this talk, I will outline a few AI/ML approaches for model-independent
cosmological inference, with an emphasis on how data-driven methods can
complement more traditional techniques. I will discuss LADDER, a deep-learning
framework for reconstructing the cosmic distance ladder directly from Type Ia
supernovae (SNIa) data using full covariance information. Following extensive
robustness tests, LADDER provides reliable predictions that enable
model-independent applications such as BAO consistency checks, calibration of
high-redshift datasets (e.g., GRBs), and the construction of mock catalogues for
future SNIa and gravitational-wave (GW) missions. These examples illustrate how
well-validated deep-learning tools can assist cosmological analyses without
assuming specific parametric forms. I will also briefly mention
Gaussian-process-based reconstruction of the Hubble parameter and its use in
examining the possible cosmological implications of future GW observations.
Taken together, these methods aim to demonstrate both the potential, and the
necessary caution, in applying AI/ML techniques to cosmology, and show how
responsible non-parametric approaches may offer fresh perspectives on several
ongoing cosmological challenges.
