Presentation #311.04D in the session “Star-forming Regions at Distant Galaxies”.
We used the 21 cm spin-flip transition of neutral hydrogen during the Epoch of Reionization as a probe to constrain the optical depth to Reionization with uncertainties in the reported measurement using unconventional methods. This is a potentially powerful method as the transition is directly related to the ionization state of the intergalactic medium. The optical depth to Reionization is a measure of the excitation coefficient and this is important to measure because it provides information on the transparency of the medium. Machine Learning is an extremely powerful tool because as models are exposed to new data, they have the flexibility to independently adapt and learn from previous computations. Recently Machine Learning has been a crucial technique used in Astrophysical Science to tackle difficult problems more easily than other analysis methods. While Machine Learning techniques are promising, generating accurate error estimates for parameters derived from them is still an important thing to address. We used this alternative analysis approach to extract information about the optical depth by using the Dropout method, an approximation to the Bayesian Model, and the full Bayesian Convolution Neural Networks model on simulated image cubes of different realizations of the Universe during the Epoch of Reionization. The goal of this talk will be to summarize the results from the past using non-Bayesian methods, summarize how the data was constructed, outline the architecture of the network that uses the Bayesian method, and discuss the results of the predictive model. This talk will be a follow-up to my work presented at AAS Spring 2020 using non-Bayesian methods to extract information and recovered the optical depth to reionization to within less than 2%.