演講資訊

摘要：

For decades astronomers have studied the distribution of gas in the interstellar medium by making 2-d dust emission maps and 3-d spectral line cubes. It is a community tradition to identify salient structures in these complex and noisy data, an undertaking affectionately known as "lob-ology". In computer science and statistics, this type of data segmentation task is a fundamental algorithmic challenge that boils down to: what is the best way to robustly cluster data? In the star formation content, a variety of algorithms have been developed to identify dense cores and filaments, however, these methods all depend on user-set parameters that are often tuned by hand and validated by-eye. In this talk I will discuss the pros and cons of supervised machine learning approaches to data segmentation and present results from a 3-d convolutional neural network model, which can accurately identify stellar feedback features in molecular line spectral cubes. I will introduce a new unsupervised approach to identify cohesive structures in hyperspectral cubes: Multiview Prototype Embedding and Clustering (MPEC). MPEC learns prototypes of spectral data for efficient sample size reduction, embeds these prototypes in a lower dimensional latent space generated by a Self-Organizing variant of UMAP, and clusters a graph representation of (learned) prototype similarity in both feature and latent spaces. I will review what such approaches teach us about star formation and discuss the prognosis for addressing outstanding star formation questions: Which gas will form stars? How does star-forming gas evolve? What is the future of observed molecular clouds?