Research map / Label-efficient and robust learning / Recognition and long-tail
Forward Compatible Training for Large-Scale Embedding Retrieval Systems
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use FCT transformation with SimCLR side-information instead of BCT | ResNet50-128 retrieval on ImageNet-1k · ImageNet-1k val | CMC top-1 (%): 46.9 to 65.0; +18.1% | Helped |
| Use FCT transformation with SimCLR side-information instead of BCT | ResNet50-512 retrieval on Places-365 · Places-365 val | CMC top-1 (%): +5.4% | Helped |
| Use FCT transformation with alternate old-model side-information instead of BCT | ResNet18-128 to ResNet50-128 retrieval · VGGFace2 val | CMC top-1 (%): +8.3% | Helped |
| Add SimCLR side-information rather than using no side-information (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 61.8 to 65.0 | Helped |
| Use autoencoder side-information rather than no side-information (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 61.8 to 62.0 | Helped |
| Use alternate old-model side-information rather than no side-information (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 61.8 to 63.5 | Helped |
| Add Mixup to alternate old-model side-information training (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 63.5 to 63.8; 0.3% | Helped |
| Train SimCLR side-information on ImageNet-1k rather than ImageNet-500 (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 65.0 to 66.5; +1.5% | Helped |
| Increase transformation capacity from 0.79M to 5.7M parameters (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 62.9 to 65.0 | Helped |
| Increase transformation capacity from 5.7M to 19.6M parameters (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 65.0 to 65.0 | No clear effect |
| Replace MSE transformation loss with KL loss (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 65.0 to 60.7 | Hurt |
| Replace MSE transformation loss with reversed KL loss (ablation) | FCT retrieval · ImageNet-1k val | CMC top-1 (%): 65.0 to 55.7 | Hurt |
Closest papers by meaning
- Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements (CVPR 2024 · similarity 0.87)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.86)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.86)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.85)
- QuARI: Query Adaptive Retrieval Improvement (NeurIPS 2025 · similarity 0.85)
- Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation (ICML 2025 · similarity 0.85)
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.