Tag: #self-supervised7 results
Nature 2026 paper introduces GluFormer, a Transformer-based foundation model for CGM data using self-supervised learning on 10M+ readings from 10k subjects. It outperforms HbA1c for 11-year diabetes/CVD risk prediction and generates realistic glucose curves across populations. Enables personalized diet/treatment forecasting from short-term data.
Black Forest Labs released Self-Flow, a self-supervised flow matching framework that enables multimodal AI models to learn representations and generation simultaneously without external teachers like CLIP. It uses Dual-Timestep Scheduling for self-distillation, achieving state-of-the-art results in images, video, and audio. The technique converges 2.8x faster than the REPA standard and scales without plateauing.
He Kaiming's team unveils GeoPT, a groundbreaking pre-training paradigm. It enables models to autonomously learn real-world physical laws. This shifts AI towards more intuitive physical understanding.
Tsinghua University's automation and astronomy teams used self-developed spatio-temporal self-supervised model Xingyan to create humanity's deepest deep-space galaxy image. This breakthrough surpasses astronomical observation depth limits. Results published in Science magazine.
Tsinghua researchers' ASTERIS AI model boosts James Webb Space Telescope's deep space imaging by 1 magnitude, equivalent to a 10m aperture. It discovers over 160 high-redshift galaxies from 2-5 billion years post-Big Bang, tripling prior findings. Published in Science, it's self-supervised and compatible across telescopes.
Analysis shows GluFormer's self-supervised framework transfers to ecology for unlabeled time-series like FLUXNET flux data. Addresses static metrics' limits, enabling early ecosystem risk prediction without annotations. Tokenization adapts Transformer to multi-scale ecological dynamics.
Post seeks methods to fine-tune sentence embedding aggregation beyond mean pooling from transformers, without labels or changing the base model. Focus on general self-supervised strategies for non-NLP datasets. Bonus interest in dimensionality reduction.