Cosmological Parameter Estimation with Sequential Linear Simulation-based Inference

However, the use of neural networks presents some disadvantages, the most significant of which is their lack of explainability. This means that most neural networks In many astrophysical applications, statistical models are treated as a'black box', where the decisions taken can be simulated forward, but their likelihood functions by the artificial intelligence in arriving at the optimized are too complex to calculate directly. Simulation-based solution are not known to researchers, which can hinder inference (SBI) [1] provides an alternative way to perform intellectual oversight [18]. This problem affects the Bayesian analysis on these models, relying solely on forward algorithms discussed above, as NRE constitutes an unsupervised simulations rather than likelihood estimates. However, learning task, where the artificial intelligence is modern cosmological models are typically expensive given unlabeled input data and allowed to discover patterns to simulate and datasets are often high-dimensional, in its distribution without guidance. This combines so traditional methods like the Approximate Bayesian with the problem of over-fitting, where the neural network Computation (ABC) [2], which scale poorly with dimensionality, may attempt to maximize the likelihood without are no longer suitable for parameter estimation.