Deep Transfer Learning for Star Cluster Classification: I. Application to the PHANGS-HST Survey. (arXiv:1909.02024v1 [astro-ph.GA])
<a href="http://arxiv.org/find/astro-ph/1/au:+Wei_W/0/1/0/all/0/1">Wei Wei</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Huerta_E/0/1/0/all/0/1">E. A. Huerta</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Whitmore_B/0/1/0/all/0/1">Bradley C. Whitmore</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Lee_J/0/1/0/all/0/1">Janice C. Lee</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Hannon_S/0/1/0/all/0/1">Stephen Hannon</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Chandar_R/0/1/0/all/0/1">Rupali Chandar</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Dale_D/0/1/0/all/0/1">Daniel A. Dale</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Larson_K/0/1/0/all/0/1">Kirsten L. Larson</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Thilker_D/0/1/0/all/0/1">David A. Thilker</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Ubeda_L/0/1/0/all/0/1">Leonardo Ubeda</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Boquien_M/0/1/0/all/0/1">M&#xe9;d&#xe9;ric Boquien</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Chevance_M/0/1/0/all/0/1">M&#xe9;lanie Chevance</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Kruijssen_J/0/1/0/all/0/1">J. M. Diederik Kruijssen</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Schruba_A/0/1/0/all/0/1">Andreas Schruba</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Blanc_G/0/1/0/all/0/1">Guillermo Blanc</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Congiu_E/0/1/0/all/0/1">Enrico Congiu</a>

We present the results of a proof-of-concept experiment which demonstrates
that deep learning can successfully be used for production-scale classification
of compact star clusters detected in HST UV-optical imaging of nearby spiral
galaxies in the PHANGS-HST survey. Given the relatively small and unbalanced
nature of existing, human-labelled star cluster datasets, we transfer the
knowledge of neural network models for real-object recognition to classify star
clusters candidates into four morphological classes. We show that human
classification is at the 66%:37%:40%:61% agreement level for the four classes
considered. Our findings indicate that deep learning algorithms achieve
76%:63%:59%:70% for a star cluster sample within 4Mpc < D <10Mpc. We tested the robustness of our deep learning algorithms to generalize to different cluster images using the first data obtained by PHANGS-HST of NGC1559, which is more distant at D = 19Mpc, and found that deep learning produces classification accuracies 73%:42%:52%:67%. We furnish evidence for the robustness of these analyses by using two different neural network models for image classification, trained multiple times from the ground up to assess the variance and stability of our results. We quantified the importance of the NUV, U, B, V and I images for morphological classification with our deep learning models, and find that the V-band is the key contributor as human classifications are based on images taken in that filter. This work lays the foundations to automate classification for these objects at scale, and the creation of a standardized dataset.

We present the results of a proof-of-concept experiment which demonstrates
that deep learning can successfully be used for production-scale classification
of compact star clusters detected in HST UV-optical imaging of nearby spiral
galaxies in the PHANGS-HST survey. Given the relatively small and unbalanced
nature of existing, human-labelled star cluster datasets, we transfer the
knowledge of neural network models for real-object recognition to classify star
clusters candidates into four morphological classes. We show that human
classification is at the 66%:37%:40%:61% agreement level for the four classes
considered. Our findings indicate that deep learning algorithms achieve
76%:63%:59%:70% for a star cluster sample within 4Mpc < D <10Mpc. We tested the
robustness of our deep learning algorithms to generalize to different cluster
images using the first data obtained by PHANGS-HST of NGC1559, which is more
distant at D = 19Mpc, and found that deep learning produces classification
accuracies 73%:42%:52%:67%. We furnish evidence for the robustness of these
analyses by using two different neural network models for image classification,
trained multiple times from the ground up to assess the variance and stability
of our results. We quantified the importance of the NUV, U, B, V and I images
for morphological classification with our deep learning models, and find that
the V-band is the key contributor as human classifications are based on images
taken in that filter. This work lays the foundations to automate classification
for these objects at scale, and the creation of a standardized dataset.

http://arxiv.org/icons/sfx.gif