General Profile: 

I graduated with an MEng in Computer Science here at Bristol. Although I enjoyed many topics throughout my degree, I found myself drawn to units involving data-driven learning and applications. My interest in these subjects led me to complete a Deep Learning based final year project on "Synthesising Human Motion Using Recurrent Neural Networks".

Research Project Summary:

Supermassive black holes are found in the centres of all massive galaxies. During the active phase of accretion of matter, black holes reveal themselves as Active Galactic Nuclei (AGN) due to the electromagnetic radiation emitted by the gas in the vicinity of the black hole. In the extreme case when the AGN outshines the host galaxy, we have a quasar, or quasi stellar object (QSO). During the AGN/QSO phase, strong winds can deposit energy onto the host galaxy, enough to suppress the formation of new stars. However, the exact mechanism of this feedback is currently under heated discussion. A major challenge in establishing a complete picture regarding the role of AGN in galaxy evolution is the comparatively short timescale of the AGN phase, making them rare objects (less than 1 in 1000 galaxies). Furthermore, the selection criteria used to identify AGN (X-rays, infrared, radio, optical) lead to galaxy samples with very little overlap. The latter is potentially connected to an evolutionary scenario of AGN.

The impact of accurate source classification is not limited to galaxy evolution and AGN studies. Many planned large facilities aiming to study dark matter and dark energy require very precise classifications for all sources detected. For example, the Euclid mission is a European Space Agency (ESA) telescope due to launch in the summer of 2022. With a budget of €1 billion, the 6-year long program will survey almost the entire extragalactic sky and will look back at the equivalent of 10 billion years in the past. The Euclid mission aims to shed light onto the dark universe (dark matter and dark energy) by pin-pointing the cosmological parameters with high accuracy. The tight constraints imposed by the aim of the mission, define the performance at each step of the data analysis. The success of the cosmological analysis, in particular weak lensing analysis, is largely based on 1) the accurate photometric redshift (distance) estimation and 2) modelling of the point-spread function of the telescope. Accurate classification is paramount for both tasks, as the ability to identify pure galaxy (without any stars, AGN/QSO) and pure star samples (without galaxies, AGN/QSO) will impact the error budget. Successfully identifying AGNs and other astronomical objects will reduce sources with catastrophically wrong redshifts which, with a project this big, is critical for both time and resources. At the same time, a pristine and large sample of AGN/QSO will have a significant impact on galaxy evolution models, particularly on the study of the coevolution of galaxies and AGN.

This project will assess the use of active learning and outlier detection methods for identifying AGN. These methods aim to address key challenges such as vastly imbalanced datasets, ambiguous class definitions which lead to labels which cannot be entirely trusted and vital characteristics that may not present themselves in all wavelengths. By incorporating the knowledge of domain experts into the labelling process, more reliable and accurate labels will be used in model training, allowing for more accurate predictions and requiring far less data. By thinking of AGN as an outlier to the other classes, rather than its own class, we can take advantage of the improved performance in detection of the other classes.

This project also includes the development of bespoke software for both data analysis and model training. By providing as full a picture of each data source as possible, this interactive platform will allow researchers to take full advantage of the produced methods and apply them to other astronomical classification problems, all in preparation for the huge amounts of new data to arrive in the coming years.