Research map / Architectures and efficiency / CNN design
Deep Hyperspherical Learning
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 standard convolutions with SphereConv and use GA-Softmax | ResNet-32 · CIFAR-10+ | accuracy: 93.26 -> 95.01 | Helped |
| Replace standard convolutions with SphereConv and use GA-Softmax | ResNet-32 · CIFAR-100 | accuracy: 72.85 -> 76.39 | Helped |
| Use SphereConv instead of original convolution (ablation) | CNN-9 with GA-Softmax · CIFAR-10, no augmentation | accuracy: 91.80 -> 92.38 | Helped |
| Use GA-Softmax instead of original softmax (ablation) | Sigmoid SphereNet (k=0.3) · CIFAR-10, CNN-9, no augmentation | accuracy: 91.08 -> 92.38 | Helped |
| Use GA-Softmax instead of A-Softmax (ablation) | Sigmoid SphereNet (k=0.3) · CIFAR-10, CNN-9, no augmentation | accuracy: 92.13 -> 92.38 | Helped |
| Learn sigmoid SphereConv parameter k independently for each filter (ablation) | CNN-9 SphereNet · CIFAR-10, no augmentation | accuracy: 91.22% -> 91.64% | Helped |
| Increase network width (ablation) | SphereNet · CIFAR-10 | accuracy | Helped |
| Remove ReLU while retaining SphereConv (ablation) | CNN-18 SphereNet · CIFAR-10, without ReLU | accuracy: 80.73 -> 86.29 | Helped |
| Add orthogonality constraints on kernel weights (ablation) | SphereNet · CIFAR-10 | accuracy | Mixed |
| Use SphereNorm instead of BatchNorm (ablation) | CNN-9 · CIFAR-10 with small mini-batches | accuracy | No clear effect |
| Use SphereNorm together with BatchNorm (ablation) | CNN-9 · CIFAR-10 with small mini-batches | accuracy | Helped |
| Rescale SphereNorm output (ablation) | CNN-9 SphereNorm · CIFAR-10 with small mini-batches | accuracy | Hurt |
Closest papers by meaning
- DeepSphere: a graph-based spherical CNN (ICLR 2020 · similarity 0.87)
- Scalable and Equivariant Spherical CNNs by Discrete-Continuous (DISCO) Convolutions (ICLR 2023 · similarity 0.86)
- O$n$ Learning Deep O($n$)-Equivariant Hyperspheres (ICML 2024 · similarity 0.84)
- Bridging Mini-Batch and Asymptotic Analysis in Contrastive Learning: From InfoNCE to Kernel-Based Losses (ICML 2024 · similarity 0.84)
- Fully Hyperbolic Convolutional Neural Networks for Computer Vision (ICLR 2024 · similarity 0.83)
- HYPO: Hyperspherical Out-Of-Distribution Generalization (ICLR 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.