FDSML, accepted at WACV 2027, lifts formula-driven pretraining Recall@1 to 74%
HirokatuKataoka · x · 2026-11-01
A Japan-based team's paper "FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss" is accepted at WACV 2027, with code and dataset released. Key ideas:
- Moves beyond casting formula-driven supervised learning (FDSL) as classification on discretized labels; instead derives relational supervision from the underlying image-generation (Iterated Function System) parameter space.
- Introduces FineFractalDB, a refined fractal dataset with higher intra-class variation, plus a Parameter-Aware Triplet (PA-Triplet) Loss that assigns positives/negatives by distances in parameter space.
- Results on ResNet-50: Recall@1 improves from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars over FDSL pretraining, with competitive classification and better robustness under corruption.
Takeaway: parameter-derived relational supervision is a stronger synthetic pretraining signal than label-based classification for transfer beyond recognition.
Related event: FDSML Paper on Formula-Driven Metric Learning Accepted by WACV 2027(3 posts)→
More from Research
- OpenAI's Astra decimates specialized robotics models, sparking scale-transfer debate — willdepue · 2026-11-01
- Give agents a sense of elapsed time and budget — models derive surprising hacks — realsohamparekh · 2026-11-01
- Interpretability finding: personas farther from the assistant share more top neurons — Sauers_ · 2026-11-01
- Free YouTube series teaches linear algebra for AI & ML, covering SVD to eigendecomposition — wufuheng · 2026-11-01
- Descartes' Demon Is Back: New Paper Shows Image Manipulation Can Fool AI Perception — PostponeIdiocracy · 2026-11-01
- New skeletal muscle single-cell eQTL resource from 287 individuals — anshulkundaje · 2026-11-01