Speaker
Description
In this talk, we will present our results in leveraging deep learning techniques for the automatic classification of solar active regions, for both the Mount Wilson and the McIntosh classification schemes. For this latter one, we consider a hierarchical multitask learning approach that mirrors the dependency structure inherent in the McIntosh system, which decomposes sunspot morphology into three components: the modified Zurich class (Z), penumbral class (p), and compactness class (c). We will present advanced model training techniques, including the teacher forcing method applied in the McIntosh classification. This method proves useful for enhancing training stability and convergence speed. It also mitigates error propagation by incorporating ground truth labels as input for subsequent tasks, with its influence gradually decreasing throughout the training process.
| Do you plan to attend in-person or online? | In-person |
|---|
