https://lelain.net/tags/self-supervised-learning/Self-Supervised LearningHugoBlox Kit (https://hugoblox.com)en-usFri, 25 Sep 2026 00:00:00 +0000https://lelain.net/media/icon.svghttps://lelain.net/tags/self-supervised-learning/https://lelain.net/publications/aurora-ssl/Fri, 25 Sep 2026 00:00:00 +0000https://lelain.net/publications/aurora-ssl/<p>Preprint posted on arXiv on September 25, 2026, and submitted to IEEE ICASSP 2027. A joint work with the Royal Belgian Institute for Space Aeronomy, based on the ASIS archive of auroral spectra acquired in Skibotn, Norway.</p> <p>This work follows the supervised study presented at , where a 1D Vision Transformer was trained on 719 expert-annotated spectra. Here, the unlabelled part of the archive (223,000 spectra) is used to pretrain the same architecture with a masked autoencoder, which raises the macro-AP from 77.8 to 88.5 on the same benchmark and shows that the learned representation encodes the emission-line ratios used by physicists, without being told about them.</p>