Research map / Label-efficient and robust learning / Recognition and long-tail
Large-Margin Softmax Loss for Convolutional Neural Networks
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 |
|---|---|---|---|
| replace softmax loss with L-Softmax (m=2,3,4) | CNN on MNIST · MNIST | error rate (%): 0.40 -> 0.32/0.31/0.31 | Helped |
| replace softmax loss with L-Softmax (m=2,3,4) | CNN on CIFAR10 (no augmentation) · CIFAR10 | error rate (%): 9.05 -> 7.73/7.66/7.58 | Helped |
| replace softmax loss with L-Softmax (m=2,3,4) | CNN on CIFAR10+ (with augmentation) · CIFAR10+ | error rate (%): 6.50 -> 6.01/5.94/5.92 | Helped |
| replace softmax loss with L-Softmax (m=2,3,4) | CNN on CIFAR100 · CIFAR100 | error rate (%): 32.74 -> 29.95/29.87/29.53 | Helped |
| replace softmax loss with L-Softmax (m=2,3,4) trained on WebFace | CNN for LFW face verification · LFW | verification accuracy (%): 96.53 -> 97.81/98.27/98.71 | Helped |
| increase margin m from 2 to 4 in L-Softmax (ablation) | CNN on CIFAR100 · CIFAR100 | error rate (%): 29.95 -> 29.53 | Helped |
| increase margin m from 2 to 4 in L-Softmax (ablation) | CNN on CIFAR10+ · CIFAR10+ | error rate (%): 6.01 -> 5.92 | Helped |
| increase margin m from 2 to 4 in L-Softmax (ablation) | MNIST · MNIST | error rate (%): 0.32 -> 0.31 | No clear effect |
| replace softmax with hinge loss (ablation) | CNN on CIFAR10 · CIFAR10 | error rate (%): 9.05 -> 9.91 | Hurt |
| add contrastive loss to softmax (ablation) | CNN on LFW · LFW | verification accuracy (%): 96.53 -> 97.31 | Helped |
Closest papers by meaning
- Learning Towards The Largest Margins (ICLR 2022 · similarity 0.92)
- Rethinking Feature Distribution for Loss Functions in Image Classification (CVPR 2018 · similarity 0.91)
- Mis-Classified Vector Guided Softmax Loss for Face Recognition (AAAI 2020 · similarity 0.89)
- Large Margin Deep Networks for Classification (NeurIPS 2018 · similarity 0.88)
- Stop-Gradient Softmax Loss for Deep Metric Learning (AAAI 2023 · similarity 0.88)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.87)
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.