Learning What's Real: Disentangling Signal and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics
A deep learning framework that leverages overlapping observations, a dual-encoder architecture, and a counterfactual generation objective to disentangle factors of variation and provide a general recipe for scientific and multi-modal self-supervised pretraining.
Pablo Mercader-Perez, C. Cuesta-Lázaro, D. Muthukrishna et al.
· arXiv.org · 0 citations