Results
Preliminary Results on classification of the Oslo Aurora THEMIS (OATH) dataset
This work was inspired by preliminary classification results on the Oslo Aurora THEMIS dataset, where our team developed a self-supervised machine learning algorithm based on the Simple Method for Contrastive Learning of Visual Representations (SimCLR) algorithm that obtained state-of-the-art classification performance. Going beyond classification performance, the visual representations learned by our model naturally clustered into several more qualitatively distinct categories of images than exist manually assigned labels, suggesting both that current labeling of auroral images is too coarse and that the distinct categories identified by the model may provide insight into the as yet not fully understood magnetospheric drivers. These early results were published in a 2021 paper at the 20th International Conference on Machine Learning and Applications (ICMLA 2021). Stay tuned for more results as the project continues to unfold!