Research map / Label-efficient and robust learning / Recognition and long-tail
Accelerated Training for Massive Classification via Dynamic Class Selection
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 |
|---|---|---|---|
| Use hashing-forest active-class selection instead of full softmax | Hynet · IJB-A | TPR at FPR 0.001: 0.860 -> 0.856 | No clear effect |
| Use hashing-forest active-class selection instead of full softmax | Hynet · Megaface & Facescrub | top-1 accuracy: 0.630 -> 0.599 | Hurt |
| Use hashing-forest active-class selection instead of random selection | Hynet · IJB-A | TPR at FPR 0.001: 0.771 -> 0.856 | Helped |
| Use hashing-forest active-class selection instead of random selection | Hynet · Megaface & Facescrub | top-1 accuracy: 0.515 -> 0.599 | Helped |
| Add adaptive allocation of active-class quota and hashing resources (ablation) | HF · Megaface & Facescrub | top-1 identification accuracy | Helped |
| Increase the number of active classes per iteration (ablation) | Random, Optimal Selection, and HF · Face identification | accuracy | Helped |
| Increase the number of hashing trees (ablation) | HF · Face identification | accuracy | Helped |
| Increase hashing trees beyond 100 (ablation) | HF · Face identification | accuracy | No clear effect |
| Vary the forest update interval from 100 to 1000 iterations (ablation) | HF · Face identification | accuracy | No clear effect |
| Increase the forest update interval from 1000 to 10000 iterations (ablation) | HF · Face identification | accuracy | Mixed |
| Raise the adaptive-allocation threshold τcp from 0.7 to 0.9 (ablation) | HF-A · Face identification | accuracy: about 2% | Helped |
| Raise τcp beyond 0.9 (ablation) | HF-A · Face identification | accuracy | No clear effect |
Closest papers by meaning
- Dynamic Class Queue for Large Scale Face Recognition in the Wild (CVPR 2021 · similarity 0.91)
- An Efficient Training Approach for Very Large Scale Face Recognition (CVPR 2022 · similarity 0.88)
- Multi-Scale Dense Networks for Resource Efficient Image Classification (ICLR 2018 · similarity 0.87)
- Virtual Fully-Connected Layer: Training a Large-Scale Face Recognition Dataset With Limited Computational Resources (CVPR 2021 · similarity 0.86)
- Hypernetworks for Dynamic Feature Selection (NeurIPS 2026 · similarity 0.85)
- Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness (ICLR 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.