Ingo Waldmann
Department of Physics and Astronomy, University College London (Londres, Royaume-Uni)
For low to mid signal to noise observations, we are prone to two sources of biases: 1) Biases in the data reduction and analysis through an inaccurate knowledge of the instrument response function; 2) Biases through prior constraints on the spectral retrieval.
Subjectivity in analysis and interpretation often leads to controversy with different teams obtaining different results, even for the same data sets. Supervised and unsupervised machine learning can provide a more objective way forward here.
Using so called blind-source separation (BSS) techniques, a form of unsupervised learning, we have shown that instrument systematics can be objectively de-trended from the science signal. The only prior assumption is that both instrument and science are statistically independent from one another. This approach has been validated on both spectroscopic as well as photometric data.
Following these approaches we also developed a new take on the spectral retrieval of extrasolar planets. Tau-REx (tau-retrieval of exoplanets) is a line-by-line, bayesian atmospheric retrieval framework. Atmospheric retrievals are often hampered by large correlated parameter spaces that depend on user-defined constraints. Tau-REx takes a different approach and does not require user input. It selects its appropriate parameter space from a huge space of possible solutions using pattern recognition software and neural networks trained to recognise spectral signatures in exoplanetary atmospheres.
In this seminar, I will give an overview of the machine learning efforts in the field of exoplanets to date, what ‘dreaming’ neural networks can tell us about atmospheric characteristics and how statistical independence can hold the key to de-trending a wide range of instrument systematics.
Amphithéâtre Henri Mineur, Institut d'Astrophysique de Paris