Research map / Label-efficient and robust learning / Recognition and long-tail
FastFill: Efficient Compatible Model Update
Florian Jaeckle, Fartash Faghri, Ali Farhadi, Oncel Tuzel, Hadi Pouransari
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FCT with FastFill’s alignment and uncertainty-based partial-backfilling policy | Compatible embedding update · ImageNet-1k | average mAP: 40.47 -> 44.84 | Improved |
| Replace FCT with FastFill’s alignment and uncertainty-based partial-backfilling policy | Compatible embedding update · Places-365 | average mAP: 16.81 -> 19.48 | Improved |
| Replace FCT with FastFill’s alignment and uncertainty-based partial-backfilling policy | Compatible embedding update · VGGFace2 | average mAP: 70.26 -> 71.57 | Improved |
| Add the discriminative objective to L2 alignment (ablation) | FastFill alignment; random ordering · ImageNet-1k | average mAP: 40.47 -> 41.95 | Improved |
| Use discriminative loss alone instead of L2 alignment (ablation) | FastFill alignment; random ordering · ImageNet-1k | average mAP: 40.47 -> 39.45 | Hurt |
| Add uncertainty estimation while keeping random backfilling and L2 alignment (ablation) | FastFill · ImageNet-1k | average mAP: 40.47 -> 40.08 | No clear effect |
| Use predicted uncertainty to order backfilling rather than random order (ablation) | FastFill with L2 + discriminative loss and uncertainty · ImageNet-1k | average mAP: 42.05 -> 44.84 | Improved |
| Use discriminative loss alone with uncertainty-based ordering instead of the combined alignment loss (ablation) | FastFill · ImageNet-1k | average mAP: 44.84 -> 27.59 | Hurt |
| Use Bayesian uncertainty instead of entropy uncertainty (ablation) | FastFill backfilling policy · Places-365 | average mAP: 18.6 -> 19.48 | Improved |
| Use Bayesian uncertainty instead of entropy-based uncertainty (ablation) | FastFill backfilling policy · ImageNet-250 to ImageNet-500, disjoint classes [500:] | average top1: 15.22 -> 17.9 | Improved |
| Use Bayesian uncertainty instead of margin-confidence uncertainty (ablation) | FastFill backfilling policy · ImageNet-250 to ImageNet-500, disjoint classes [500:] | average top1: 15.17 -> 17.9 | Improved |
| Use Bayesian uncertainty instead of least-confidence uncertainty (ablation) | FastFill backfilling policy · ImageNet-250 to ImageNet-500, disjoint classes [500:] | average top1: 15.18 -> 17.9 | Improved |
Later papers that compared with this method
1 later accepted paper reports FastFill as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- $\boldsymbol{\lambda}$-Orthogonality Regularization for Compatible Representation Learning (NeurIPS 2025 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements (CVPR 2024 · similarity 0.87)
- QuARI: Query Adaptive Retrieval Improvement (NeurIPS 2025 · similarity 0.85)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.85)
- LOCORE: Image Re-ranking with Long-Context Sequence Modeling (CVPR 2025 · similarity 0.84)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.84)
- MOFI: Learning Image Representations from Noisy Entity Annotated Images (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
