Thursday

Astronomy is simultaneously plagued by the high dimensionality of its noisy data, bias induced from sample incompleteness and assumptions contained within parametric models.
There is not one tool that attempts to mitigate these issues. Extreme deconvolution and PyGMMIS infer an underlying Gaussian mixture model for the entire incomplete and noisy dataset but cannot deal with non-gaussian uncertainties and cannot return the full posterior distribution.
Here we introduce an expanded version of extreme deconvolution whereby we train a neural network to approximate non-Gaussian log-likelihoods and fit the multimodal distribution with nested sampling.
This gives us access to the full posterior distribution of the underlying distribution even when the data is incomplete and their uncertainties are not Gaussian.
We describe how we used neural-network enhanced extreme deconvolution to fit the LOFAR-luminosity stellar-mass distribution in order to quantify the slope of the radio-synchrotron luminosity correlation.