Astrofísica

When we train a machine learning model from data annotated by the crowd, there is uncertainty about the annotations for several reasons. This is known as “learning from disagreement”. In this talk I will give a brief introduction to learning from disagreement, including the different approaches. Then I will share an experimental analysis we have done on developing deep learning models from crowd annotations for galaxy morphological classification. In particular, we analyze the results considering different dimensions: (I) training with different levels of agreement between annotators, (ii) training only the last layer or all the layers of a deep neural network, (iii) incorporating the hierarchy in the model in a post-processing stage, (iv) varying the amount of training data, (v) training in several stages (curriculum learning). We present results from the SLOAN galaxy survey based on Galaxy Zoo 1, and initial results with Galaxy Zoo 2, including a proposal for hierarchical agreement.

tbd

tbd

tbd

tbd

tbd

tbd

tbd

tbd

tbd

tbd

tbd

tbd