| Abstract : | The Epoch of Reionization and Cosmic Dawn marks the two essential epochs in the structure formation in the first billion years of the Universe. The properties of the Intergalactic Medium (IGM) during these two epochs have not yet been observationally constrained. Observation of redshifted 21 cm signal from these epochs is a promising probe into this first billion years. This forms a key science goal of ongoing/upcoming experiments like the EDGES, SARAS, HERA, MWA, LOFAR, and SKA. However, these experiments suffer from systematics and are heavily dependent on the accuracy of foreground removal. Hence, it is crucial to understand the effect of each corrupting term in the detection of the signal through non-parametric techniques like Machine Learning or Bayesian Statistics. In the absence of any observational constraint, the signal parameter space is overwhelmingly large. Hence, it is critical to sample it efficiently by creating a representative subsample as our input dataset for 21 cm signal extraction training. Often, such non-parametric approaches do not span the entire parameter space, which in turn creates a bias in the resulting inferences. In this work, we are constructing the data set with various realizations of the Global 21 cm signal by sampling the whole parameter space. To reduce the computational cost of training the Artificial Neural Networks (ANNs) model and achieve more accuracy in parameter extraction from the Global Signal, we are selecting appropriate samples from the parameter space by applying different sampling methods. We are exploring three techniques: Random, Stratified, and quasi Monte-Carlo Sampling and comparing their results. The initial finding is that the Neural network trained on a more diverse signal set extracts the signal parameters with the best accuracy when tested on an unknown signal.
|
