metric-learning

MABLE combines masked reconstruction with cosine-similarity losses to learn node and graph embeddings from large heterogeneous graphs, demonstrated on geospatial mineral-exploration data. It unifies masked autoencoding and metric learning in a self-supervised framework without requiring labeled data.

This paper introduces the Kalman Prototypical Network (KPN), a few-shot learning framework for fault detection in combined-cycle gas turbines. KPN models class prototypes as latent stochastic states to reduce variance and outperforms conventional methods on simulated leak detection tasks.

SAGA framework uses frozen multimodal large language models to provide attribute-aware supervision for vision encoders via Group Relative Policy Optimization, improving zero-shot image retrieval by 3–6 points on fine-grained benchmarks.