Research map / Label-efficient and robust learning / Recognition and long-tail
Pairwise Similarity Learning is SimPLE
Yandong Wen, Weiyang Liu, Yao Feng, Bhiksha Raj, Rita Singh, Adrian Weller, Michael J. Black, Bernhard Schölkopf
Read the paperCheck 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 |
|---|---|---|---|
| use reverse-direction hard pair mining (r=3) instead of r=1 (vanilla BCE) (ablation) | SimPLE · Face validation set | EER (%): 3.72 -> 3.23 | Improved |
| use generalized inner product with angular bias instead of cosine similarity (ablation) | SimPLE · Face validation set | EER (%): 4.81 -> 3.23 | Improved |
| use generalized inner product instead of cosine in SimPLE (ablation) | SimPLE SFNet-20 · IJB-B 1:1 TAR@FAR=1e-5 | TAR (%): 63.09 -> 84.51 | Improved |
| apply generalized inner product (angular bias) to proxy-based cross-entropy (ablation) | proxy-based CE loss · Setting A validation | EER (%): 6.13 (CE) vs 5.99 to 9.83 across b_theta | No clear effect |
| add angular margin with generalized inner product to CosFace | CosFace · Setting A | convergence | Hurt |
| drop proxies, angular similarity and margin (SimPLE vs SphereFace2) in small setting | SimPLE · IJB-B 1:1 TAR@FAR=1e-5 (SFNet-20) | TAR (%): 77.13 -> 84.51 | Improved |
| use SimPLE with a stronger backbone and larger data (IResNet-100, MS1MV2) versus AdaFace | SimPLE · IJB-B 1:1 TAR@FAR=1e-4 | TAR (%): 94.78 vs AdaFace 95.67 | Mixed |
| replace softmax-based losses with SimPLE for speaker verification | ResNet-34 · VoxCeleb1 | EER (%): 2.11 -> 1.85 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 4 on most shared results.
- AdaFace — AdaFace: Quality Adaptive Margin for Face Recognition (CVPR 2022) · better on 2 of 3 shared results
- SphereFace2 — SphereFace2: Binary Classification is All You Need for Deep Face Recognition (ICLR 2022) · better on 3 of 3 shared results
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 2 of 3 shared results
- BroadFace — BroadFace: Looking at Tens of Thousands of People at Once for Face Recognition (ECCV 2020) · better on 0 of 1 shared result
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
- Large Scale Similarity Learning Using Similar Pairs for Person Verification (AAAI 2016 · similarity 0.88)
- Circle Loss: A Unified Perspective of Pair Similarity Optimization (CVPR 2020 · similarity 0.86)
- Pairwise Relational Networks for Face Recognition (ECCV 2018 · similarity 0.85)
- Similarity Learning With Spatial Constraints for Person Re-Identification (CVPR 2016 · similarity 0.85)
- Stabilizing Sample Similarity in Representation via Mitigating Random Consistency (ICML 2025 · similarity 0.84)
- Int*-Match: Balancing Intra-Class Compactness and Inter-Class Discrepancy for Semi-Supervised Speaker Recognition (AAAI 2025 · similarity 0.83)
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.
