Active Learning for Supernova photometric classification

The idea of this project is to use Active Learning (AL) to deal with the dearth of spectroscopic data that can be used as a training set for supervised statistical learning algorithms. We plan to ameliorate this issue by devising an algorithm that could effectively infer spectroscopic properties from photometric data, which is comparatively plentiful. The project will be developed within the Large Synoptic Survey Telescope (LSST), which will guarantee the long term impact of its results.

AL is generally used when it is too expensive/time consuming to label/visually classify instances. So, by developing an AL algorithm that can effectively and realistically classify supernovae, this project will also be an example of how astronomical data can be used to develop new statistical learning algorithms. This project is perfectly suited for the ‘big data overload’ scenario encountered by large surveys like the LSST.

Active anomaly detection for time domain discoveries

This project featured the first application of adaptive machine learning to identify anomalous observations in non-periodic astronomical light curves. In this approach objects that were deemed highly informative were chosen to be labelled using an AL framework in which the model would evolve in every subsequent training iteration, better informing the learning process. In the case of anomaly detection, the goal is to be able to detect more ‘real’ anomalies to present to an expert (human) classifier by marginally modifying the decision boundary of the chosen algorithm, an isolation forest (IF) in this case. When this Active Anomaly Detection (AAD) algorithm was applied to the Open Supernova Catalog, it was found that the AAD was able to identify ∼ 80%
more real anomalies as compared to the static IF from the previous installment of this work (Pruzhinskaya et. al. 2019). An initial report of this work has been submitted to A& A  (Ishida 2019, We are currently expanding this project to include different unsupervised learning algorithms such as Local Outlier Factor, One-class SVM and Gaussian Mixture Model.

Automatic morphological classification of Galaxies with machine learning techniques

Our enquiry uses proprietary data for ~8000 galaxies from the GAMA (Galaxy and Mass Assembly) Survey and employs it through the machine learning techniques of Classification trees, Classification trees with Random Forests, Support Vector Machines and Neural Networks to provide a method of classifying galaxies according to their morphological types with a relatively small sample set.

(In collaboration with Univ-Prof. Dr. Markus Haltmeier, Dr. Sergiy Pereverzyev (Department of Applied Mathematics, University of Innsbruck) and Dr. Lee Kelvin (Astrophysics Research Institute, Liverpool John Moores University)).

Unveiling hidden structure within and around early-type Galaxies

We aim to study merger remnants in Early-type galaxies for a sample of 111 galaxies from Ruiz (2015). These structures manifest in the form of tidal tails, plumes, shells, umbrellas etc, which are an echo of past merging activity and are therefore cosmologically interesting to study. In particular, we aim to calculate the light and stellar mass fraction present in these structures as a function of the local environment and quantify their morphologies. Our results will ultimately allow us to place constraints on galaxy formation mechanisms, providing observational evidence for a large, statistically meaningful sample of galaxies. We have finalised our methodology and are now applying it to all the galaxies in our sample.

(In collaboration with Dr. Lee Kelvin and Dr. Ignacio Trujillo (Instituto de Astrofisica de Canarias, La Laguna, Tenerife)).

Stellar content of Planck Sunyaev-Zeldovich Galaxy clusters

We aim to measure the stellar mass component of galaxy clusters selected via the Sunyaev-Zeldovich effect (SZE) from the Planck all-sky maps. Our primary sample consists of the 813 clusters with redshifts listed in the Planck Data Release 1 (DR1) catalogue, ranging from z = 0.0111 – 0.972. The photometric catalogues for near-IR (Ks-band) are taken from the Two Micron All Sky Survey (2MASS).  We find that the scaling relation between the Planck SZ signal and the stellar mass that we have obtained is in agreement with the scaling relation between the Planck SZ signal and the mass estimate from the X-ray calibrated SZ proxy given in the Planck DR1 catalogue. While the scaling relation that we obtained appears to be viable, our sample has been reduced greatly because of the low detection limit of 2MASS. We are on the look out for other all-sky (because of the sparseness of the Planck data) near-IR surveys with a deeper detection limit which might help us increase and refine our sample.