FDSML pretrains without real images, lifting image retrieval Recall@1 by 22.2 points
HirokatuKataoka · x · 2026-10-31
Accepted at WACV 2027, FDSML (Formula-Driven Supervised Metric Learning) extends formula-driven pre-training from classification to metric learning, using FractalDB generation parameters as supervision. It pairs a refined fractal dataset (FineFractalDB) with a Parameter-Aware Triplet Loss that assigns positives/negatives by distance in the Iterated Function System parameter space. Without any real images, ResNet-50 Recall@1 rises from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars, while staying competitive on classification and improving corruption robustness.
More from Research
- Can LLMs turn surprising observations into research questions and discover physical laws? — rohanpaul_ai · 2026-11-01
- LongChang benchmark: GPT-6-Astra hits 16.7% discovery rate, other 11 LLMs score zero — rohanpaul_ai · 2026-11-01
- 747 AI-agent ideas reveal a scaling law for idea descent: +17% per doubling — DimitrisPapail · 2026-11-01
- Xiaomi's A-RSI lifts 11-benchmark average from 57.77 to 61.05 via adaptive task refresh — mark_k · 2026-11-01
- Berkeley/Stanford paper: agents gain up to 39% by wiping memory and restarting from best result — alexcovo_eth · 2026-11-01
- UniT framework trains unified multimodal models to plan, inspect, and refine iteratively — burkov · 2026-11-01