Research map / Label-efficient and robust learning / Recognition and long-tail
Hyperspherical Classification with Dynamic Label-to-Prototype Assignment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a hyperspherical classifier with fixed, uniformly distributed prototypes whose label-to-prototype assignment is optimized during training via Hungarian bipartite matching alternating with gradient descent.
Problem
Non-parametric fixed-prototype classifiers assign each class to a static prototype, which fails to capture inter-class relationships and constrains metric space utilization, especially when metric dimension is much smaller than the number of classes.
Approach
Fix equidistributed hyperspherical prototypes (via a Gaussian potential uniformity loss) and treat the label-to-prototype mapping as an optimization variable, updated each epoch by Hungarian matching between momentum-averaged class features and prototypes, while the backbone is trained with a cosine regression loss to the assigned prototype.
Key results
- On CIFAR-100 with ResNet-32 the method reaches 66.22% accuracy, 1.22% above the best HPN result, and on ImageNet-200 it reaches 46.85%, outperforming HPN and ETF.
- With metric dimension d = c/2, it outperforms ETF by 0.87% on CIFAR-100 and 9.15% on ImageNet-200 (Table 2).
- On ImageNet-1K with ResNet-101 it reports 79.63% vs 77.80% for DNC; on ResNet-50 it reports 77.47% vs 76.51% for PSC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dynamic label-to-prototype assignment instead of fixed assignment | Hyperspherical classification · CIFAR-100 | accuracy: +0.88 | Helped |
| Use dynamic label-to-prototype assignment instead of fixed assignment | Hyperspherical classification · ImageNet-200 | accuracy: +2.15 | Helped |
| Use dynamic label-to-prototype assignment | Hyperspherical classification · ImageNet-1K | accuracy: 76.51 -> 77.47 | Helped |
| Use dynamic label-to-prototype assignment | Hyperspherical classification · CIFAR-100 long-tail, imbalance factor 0.01 | accuracy: 45.3 -> 44.9 | Hurt |
| Use dynamic label-to-prototype assignment | Hyperspherical classification · ImageNet-LT | accuracy: 44.7 -> 44.6 | No clear effect |
| Update label-to-prototype assignment at τ′=10 rather than τ′=1 (ablation) | Proposed method · CIFAR-100 | accuracy | Helped |
| Replace equidistributed prototypes with pre-trained softmax prototypes (ablation) | Proposed method · CIFAR-100, d=64 | accuracy: 65.94 -> 64.71 | Hurt |
| Add prototype-uniformity regularization to PSC (ablation) | PSC · CIFAR-100, d=64 | accuracy: 63.23 -> 64.53 | Helped |
| Use dynamic prototype matching with pre-trained softmax prototypes (ablation) | Proposed method · CIFAR-100, d=64 | accuracy: 63.23 -> 64.71 | Helped |
| Use d=512 instead of d=4096 (ablation) | Proposed method on ImageNet-LT · ImageNet-LT | accuracy: 44.6 -> 44.9 | Helped |
Surprising findings
- Best performance occurs when metric dimension is smaller than the number of classes (d < c); as d approaches c the optimized prototypes converge to ETF and the assignment becomes nearly constant.
- Random zero-mean prototypes achieve optimal average pairwise angular distance, showing that APAD alone does not guarantee uniformity.
Limitations
- Assignment update cost grows as O(c^3) with the number of classes, though reported as negligible up to 10,000 classes.
- Results on some settings are mixed; e.g., the method is not uniformly best in long-tail STL-10 and SVHN results and the d=c ImageNet-LT setting is close to baselines.
- Prototype uniformity regularizer needs tuning when used as a regularizer on PSC (lambda sensitivity).
Experimental setup
Datasets: CIFAR-100 ImageNet-200 ImageNet-1K CIFAR-10 SVHN STL-10 ImageNet-LT
Benchmarks: CIFAR-100 ImageNet-200 ImageNet-1K CIFAR-10 SVHN STL-10 ImageNet-LT
Metrics: classification accuracy average pairwise cosine similarity
Baselines: PSC ETF HPN Word2Vec prototypes DNC Kasarla et al. Goto et al.
Builds on: Equiangular Tight Frame classifier Hyperspherical Prototype Networks Gaussian potential uniformity loss (Wang et al. 2020) Hungarian algorithm
Closest papers by meaning
- Hyperspherical Prototype Networks (NeurIPS 2019 · similarity 0.88)
- Learning with Noisy Labels Using Hyperspherical Margin Weighting (AAAI 2024 · similarity 0.85)
- GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · similarity 0.84)
- Prototype-based Causal Intervention for Multi-Label Image Classification (CVPR 2026 · similarity 0.83)
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.83)
- Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation (AAAI 2024 · 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.