FDSML: Supervised Metric Learning for Parameter-Aware Vision Pre-training
October 28, 2026
FDSML enables vision pre-training without real images by using parameter-aware triplet loss on fractal datasets. The method derives relational supervision from the underlying Iterated Function System parameter space, allowing embeddings to respect relative similarity orderings beyond discrete class labels.
HOW THIS AFFECTS YOU
●
researcherYou can train vision models using synthetic, parameter-driven data rather than relying solely on real-world image datasets.