Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Backward-Compatible Representation Learning
Yantao Shen, Yuanjun Xiong, Wei Xia, Stefano Soatto
Check your idea against 137,217 papers
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 |
|---|---|---|---|
| Add the old classifier's influence loss during training | New face embedding model · IJB-C 1:1 verification | TAR (%)@FAR=$10^{-4}$: 77.86 -> 80.25 | Improved |
| Add the old classifier's influence loss during training | New face embedding model · IJB-C 1:N search | TPIR (%)@FPIR=$10^{-2}$: 59.34 -> 67.23 | Improved |
| Replace basic BCT's old-data influence loss with knowledge distillation on new classes (ablation) | BCT · IJB-C 1:N search | TPIR (%)@FPIR=$10^{-2}$: 67.23 -> 69.02 | Improved |
| Use synthesized classifier weights for new classes in the influence loss (ablation) | BCT · IJB-C 1:N search | TPIR (%)@FPIR=$10^{-2}$: 67.23 -> 70.70 | Improved |
| Replace the old-data influence-loss setting with knowledge distillation for new classes (ablation) | BCT · IJB-C 1:1 verification | TAR (%)@FAR=$10^{-4}$: 80.25 -> 80.34 | Improved |
| Use synthesized classifier weights for new classes in the influence loss (ablation) | BCT · IJB-C 1:1 verification | TAR (%)@FAR=$10^{-4}$: 80.25 -> 80.59 | Improved |
| Replace BCT with an L2 distance penalty between new and old embeddings (ablation) | New face embedding model · IJB-C 1:N search | TPIR (%)@FPIR=$10^{-2}$: 67.23 -> 0.50 | Hurt |
| Train with Learning without Forgetting instead of BCT (ablation) | New face embedding model · IJB-C 1:N search | TPIR (%)@FPIR=$10^{-2}$: 67.23 -> 59.27 | Hurt |
| Use ResNet-152 instead of ResNet-101 with BCT (ablation) | BCT · IJB-C 1:N retrieval | TPIR (%)@FPIR=$10^{-2}$: 59.34 -> 68.71 | Improved |
| Increase feature dimension to 256 while using ResNet-152 and BCT (ablation) | BCT · IJB-C 1:N retrieval | TPIR (%)@FPIR=$10^{-2}$: 68.71 -> 69.45 | Improved |
| Add ReLU after the new embedding (ablation) | BCT · IJB-C 1:N retrieval | TPIR (%)@FPIR=$10^{-2}$: 68.71 -> 17.73 | Hurt |
| Use Norm-Softmax for the old model and Cosine Margin for the new model (ablation) | BCT · IJB-C 1:N retrieval | TPIR (%)@FPIR=$10^{-2}$: 64.32 -> 71.16 | Improved |
Closest papers by meaning
- Forward Compatible Training for Large-Scale Embedding Retrieval Systems (CVPR 2022 · similarity 0.92)
- Learning Compatible Embeddings (ICCV 2021 · similarity 0.88)
- Boundary-Aware Backward-Compatible Representation via Adversarial Learning in Image Retrieval (CVPR 2023 · similarity 0.87)
- FastFill: Efficient Compatible Model Update (ICLR 2023 · similarity 0.86)
- Learning Along the Arrow of Time: Hyperbolic Geometry for Backward-Compatible Representation Learning (ICML 2025 · similarity 0.86)
- Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement (CVPR 2025 · 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.
