Research map / Label-efficient and robust learning / Self-supervised and contrastive
VNE: An Effective Method for Improving Deep Representation by Manipulating Eigenvalue Distribution
Jaeill Kim, Suhyun Kang, Duhun Hwang, Jungwook Shin, Wonjong Rhee
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 |
|---|---|---|---|
| Add VNE− to ERM (ablation) | ERM · PACS, VLCS, OfficeHome, TerraIncognita | Avg. accuracy: 1.7; 1.4; 1.0; 5.2 | Improved |
| Add VNE− to SWAD (ablation) | SWAD · PACS, VLCS, OfficeHome, TerraIncognita | Avg. accuracy: 0.1; 0.3; 0.9; 0.8 | Improved |
| Add VNE+ to ERM (ablation) | ERM · PACS, VLCS, OfficeHome, TerraIncognita | Avg. accuracy: -2.9; 0.1; -0.5; -1.1 | Mixed |
| Add VNE+ to SWAD (ablation) | SWAD · PACS, VLCS, OfficeHome, TerraIncognita | Avg. accuracy: -2.1; -0.2; 0.3; -4.8 | Mixed |
| Add VNE+ to MAML (ablation) | MAML · mini-ImageNet, 5-way 1-shot and 5-shot | Avg. accuracy (%): +1.28; +1.83 | Improved |
| Add VNE+ to ANIL, BOIL, Baseline, Baseline++, and ProtoNet (ablation) | Meta-learning algorithms · mini-ImageNet, 5-way 1-shot and 5-shot | Avg. accuracy (%): 1.50; 1.34; 1.62; 2.22; 3.65; 1.92; 1.06; 3.32; 0.82; 2.48 | Improved |
| Add VNE− to meta-learning algorithms (ablation) | MAML, ANIL, BOIL, Baseline, Baseline++, ProtoNet · mini-ImageNet, 5-way 1-shot and 5-shot | Avg. accuracy (%): -2.02; -2.02; -1.30; -1.36; -1.19; -1.12; -14.98; -14.50; -18.43; -5.45; -2.10 | Hurt |
| Use I-VNE+ | Self-supervised learning · CIFAR-10 | Top-1: 93.9 -> 94.3 | Improved |
| Use I-VNE+ | Self-supervised learning · ImageNet-100 | Top-1: 84.5 -> 86.3 | Improved |
| Use I-VNE+ pretraining | ResNet-50 transfer learning · COCO detection and instance segmentation | AP / AP mask: 39.3 -> 41.2 AP; 34.4 -> 35.7 AP mask | Improved |
| Add VNE+ to GAN training (ablation) | DCGAN, PD-GAN, LOGAN, WGAN-GP, SAGAN, SNGAN, BigGAN · CIFAR-10 | FID: 42.55 -> 35.44; 28.02 -> 23.54; 18.88 -> 17.17; 24.62 -> 23.31; 9.55 -> 8.91; 9.97 -> 9.41; 5.34 -> 5.18 | Improved |
| Regularize with Frobenius norm rather than VNE (ablation) | ResNet-18 supervised training · CIFAR-10, STL-10, CIFAR-100 | Dead units: 447; 365; 325 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Barlow Twins — Barlow Twins: Self-Supervised Learning via Redundancy Reduction (ICML 2021) · better on 1 of 1 shared result
- SWAD — SWAD: Domain Generalization by Seeking Flat Minima (NeurIPS 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Variational Autoencoders Pursue PCA Directions (by Accident) (CVPR 2019 · similarity 0.85)
- Information-theoretic Generalization Analysis for VQ-VAEs: A Role of Latent Variables (NeurIPS 2025 · similarity 0.85)
- $\alpha$TC-VAE: On the relationship between Disentanglement and Diversity (ICLR 2024 · similarity 0.84)
- How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning (ICLR 2025 · similarity 0.84)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.84)
- Generalization Guarantees for Representation Learning via Data-Dependent Gaussian Mixture Priors (ICLR 2025 · 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.
