Research map / Optimization / Implicit bias and generalization
Stop-Gradient Softmax Loss for Deep Metric Learning
Read the paperCheck 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 |
|---|---|---|---|
| add SGSL to Softmax | ResNet-50 retrieval model · CUB-200-2011 | Recall@1: 74.2 -> 75.9 | Helped |
| add SGSL to Softmax | ResNet-50 retrieval model · CARS196 | Recall@1: 91.1 -> 94.7 | Helped |
| add SGSL to Softmax | ResNet-50 retrieval model · Stanford Online Products | Recall@1: 81.4 -> 83.1 | Helped |
| add SGSL to Softmax | ResNet-50 retrieval model · In-shop Clothes | Recall@1: 90.9 -> 93.0 | Helped |
| add RBR to Softmax (ablation) | ResNet-50 on CARS196 · CARS196 | Recall@1: 91.9 -> 92.4 | Helped |
| add SGSL with γ = 30 to Softmax + RBR (ablation) | ResNet-50 · CARS196 | Recall@1: 92.4 -> 94.7 | Helped |
| increase SGSL γ from 1 to 10 (ablation) | Softmax + RBR · CARS196 | Recall@1: 94.1 -> 94.6 | Helped |
| increase SGSL γ from 10 to 30 (ablation) | Softmax + RBR · CARS196 | Recall@1: 94.6 -> 94.7 | No clear effect |
| increase SGSL γ from 30 to 50 (ablation) | Softmax + RBR · CARS196 | Recall@1: 94.7 -> 94.5 | Hurt |
| increase SGSL γ from 30 to 70 (ablation) | Softmax + RBR · CARS196 | Recall@1: 94.7 -> 94.4 | Hurt |
| set SGSL loss weight β = 0.1 instead of β = 1.0 (ablation) | Softmax + SGSL · CARS196 | Recall@1: 94.7 -> 93.2 | Hurt |
| set SGSL loss weight β = 0.5 instead of β = 1.0 (ablation) | Softmax + SGSL · CARS196 | Recall@1: 94.7 -> 94.1 | Hurt |
Closest papers by meaning
- Large-Margin Softmax Loss for Convolutional Neural Networks (ICML 2016 · similarity 0.88)
- Ranked List Loss for Deep Metric Learning (CVPR 2019 · similarity 0.87)
- Towards Optimal Fine Grained Retrieval via Decorrelated Centralized Loss with Normalize-Scale Layer (AAAI 2019 · similarity 0.87)
- Improved Deep Metric Learning with Multi-class N-pair Loss Objective (NeurIPS 2016 · similarity 0.87)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · similarity 0.86)
- Mean Field Theory in Deep Metric Learning (ICLR 2024 · 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.