Research map / Architectures and efficiency / RNNs and normalization
Spherization Layer: Representation Using Only Angles
Hoyong Kim, Kangil Kim
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 |
|---|---|---|---|
| Replace last FC layer with spherization layer | VGG-11 · CIFAR100 | test accuracy: 68.42 -> 69.03 | Improved |
| Replace last FC layer with spherization layer | VGG-11 · CIFAR10 | test accuracy: 92.38 -> 92.49 | No clear effect |
| Replace last FC layer with spherization layer | VGG-11 · F-MNIST | test accuracy: 94.36 -> 94.34 | No clear effect |
| Replace last FC layer with spherization layer | SimpleFNN · MNIST | test accuracy: 98.58 -> 98.65 | No clear effect |
| Vary VGG-11 width from 16/32/64/128 to 256/512/1024/1024 with spherization | VGG-11 · CIFAR100 | test accuracy: 60.17 -> 60.40 at smallest width; 71.43 -> 71.94 at largest | Improved |
| Use spherization on deeper VGG-19 | VGG-19 · CIFAR100 | train accuracy: 97.95 -> 99.30 | Improved |
| Use spherization layer instead of direct projection with sigmoid (ablation) | CNN-9 · CIFAR10 | test accuracy: 79.03 -> 92.38 | Improved |
| Use spherization instead of cosine projection (ablation) | CNN-9 · CIFAR10 | test accuracy: 76.15 -> 92.38 | Improved |
| Use spherization instead of CW-Softmax indirect projection (ablation) | CNN-9 · CIFAR10 | error ratio among overlapping samples: 1.14 -> 0.00 | Improved |
| Use spherization instead of SW-Softmax (ablation) | CNN-9 · CIFAR10 | test accuracy: 91.51 -> 92.38 | Improved |
| Add spherization layer to BERT and RoBERTa | BERT/RoBERTa · Word analogy | average accuracy: RoBERTa 32.4 -> 33.7; BERT 33.6 -> 34.2 and 33.2 -> 33.1 | Mixed |
| Add spherization layer with Euclidean distance | ProtoNet ConvNet · Mini-ImageNet | test accuracy: 50.29 -> 43.41 | Hurt |
Closest papers by meaning
- Angular Networks: Low-Bit Learning from Randomized Similarity Estimators (NeurIPS 2026 · similarity 0.86)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.84)
- How Not to Stitch Representations to Measure Similarity: Task Loss Matching Versus Direct Matching (AAAI 2025 · similarity 0.82)
- Investigating the Benefits of Projection Head for Representation Learning (ICLR 2024 · similarity 0.81)
- Geometry-aware training of factorized layers in tensor Tucker format (NeurIPS 2024 · similarity 0.81)
- Decoupling Semantic Similarity from Spatial Alignment for Neural Networks. (NeurIPS 2024 · similarity 0.81)
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.
