FDSML
A synthetic-data pretraining method that generates parameters from mathematical formulas. It shifts the supervisory signal from discrete class labels to relational supervision over a continuous parameter space. Its core components include the FineFractalDB dataset and the PA-Triplet Loss.
Timeline (last 90 days)
FDSML (Formula-Driven Supervised Metric Learning) shifts the supervision signal from discrete class labels to image generation parameters, constructing relational supervision through a parameter-aware triplet loss (PA-Triplet Loss).
FDSML consists of two key components: the FineFractalDB dataset and the PA-Triplet Loss.
With a ResNet-50 backbone, the FDSML-pretrained model improved Recall@1 on the Stanford Online Products dataset from 65.6% to 74.0%.
FDSML remains competitive on downstream classification tasks and exhibits better robustness under various corruption settings.
All Facts (4)
FDSML (Formula-Driven Supervised Metric Learning) shifts the supervision signal from discrete class labels to image generation parameters, constructing relational supervision through a parameter-aware triplet loss (PA-Triplet Loss).
50%UnverifiedFDSML consists of two key components: the FineFractalDB dataset and the PA-Triplet Loss.
50%UnverifiedWith a ResNet-50 backbone, the FDSML-pretrained model improved Recall@1 on the Stanford Online Products dataset from 65.6% to 74.0%.
50%UnverifiedFDSML remains competitive on downstream classification tasks and exhibits better robustness under various corruption settings.
50%