Erica Hopkins (USA)
erica.hopkins @ h-its.org
A study of dimensionality reduction techniques in astrophysics
Astronomy has a long history of developing technologies which then go on to aid other fields of research and society as a whole. One potential contribution astronomy can make to the future is in the field of machine learning as we have massive amounts of different types
of data to work with.
I plan to focus my thesis on dimensionality reduction techniques. Dimensionality reduction techniques along with other unsupervised methods are a key aspect of machine learning when making large data sets accessible. Data is often very complex, existing in a high dimensional space which is hard to work with and too complex to fully understand. For example, in my master’s thesis I used a self organizing map (SOM) to reduce over 900k images, each in a high dimensional space, down to an easy to visualize and understand two
dimensional space which represents the common morphologies that appear in the 900k images. This turned a complex, nearly unexplorable data set into something which can be much more easily understood. Besides data-sets with spatial correlations, spectral and temporal pattern recognition and analysis are challenges to work on.
In astronomy, common types of measurements are images, spectra, and time series, all of which exist in a high dimensional feature space. I plan to take a broad astronomical approach and look into techniques agnostic of what type of data they are best applied to. My plan is to explore novel techniques and start building a dimensionality reduction toolkit, which can be used by other astronomers. It can be used to aid in understanding data and finding the needle in the haystack when looking for interesting scientific objects.
My personal interest is the morphological evolution and distribution of galaxies in a cosmological context. With the developed tools and methods, I hope to provide new insights for the astronomical community.
Supervisor: Kai Polsterer (HITS)