Research map / Label-efficient and robust learning / Self-supervised and contrastive
Neural Eigenfunctions are Structured Representation Learners
Zhijie Deng, Jiaxin Shi, Hao Zhang, Peng Cui, Cewu Lu, Jun Zhu
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 |
|---|---|---|---|
| use Neural Eigenmap instead of Barlow Twins | SSL pretraining · ImageNet linear probe, 100 epochs | top-1 accuracy: 66.2 -> 68.4 | Improved |
| increase batch size from 256 to 2048 (ablation) | Neural Eigenmap · ImageNet linear probe | top-1 accuracy: 60.5 -> 66.8 | Improved |
| increase batch size from 1024 to 2048 (ablation) | Barlow Twins · ImageNet linear probe | top-1 accuracy: 61.8 -> 60.4 | Hurt |
| increase projector dimension from 2048 to 8192 (ablation) | Neural Eigenmap · ImageNet linear probe | top-1 accuracy: 66.8 -> 68.4 | Improved |
| increase projector dimension to 8192 (ablation) | SCL · ImageNet linear probe | top-1 accuracy: 66.5 -> NaN | Hurt |
| add stop-gradient ordered eigenfunction learning vs SCL+PCA for short codes | Neural Eigenmap · NUS-WIDE retrieval, code length 4 | mAP: 0.4845 -> 0.6706 | Improved |
| use Neural Eigenmap instead of SCL+PCA for long codes | Neural Eigenmap · COCO retrieval, code length 32 | mAP: 0.6902 -> 0.6832 | Hurt |
| train longer from 100 to 400 epochs (ablation) | Neural Eigenmap · ImageNet linear probe | top-1 accuracy: 68.4 -> 71.5 | Improved |
| use Neural Eigenmap instead of Barlow Twins for transfer | Mask R-CNN C4 · COCO detection | AP: 39.2 -> 39.9 | Improved |
| use Neural Eigenmap instead of GCN with 10% labels | node classification · OGBN-Products | test accuracy: 73.14 -> 75.78 | Improved |
| use Neural Eigenmap instead of Laplacian Eigenmap + MLP | node classification · OGBN-Products 100% labels | test accuracy: 64.21 -> 78.33 | Improved |
| remove MLP projector or BN in hidden layer (ablation) | Neural Eigenmap · ImageNet | convergence | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Barlow Twins — Barlow Twins: Self-Supervised Learning via Redundancy Reduction (ICML 2021) · better on 9 of 9 shared results
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
- NeuralEF: Deconstructing Kernels by Deep Neural Networks (ICML 2022 · similarity 0.89)
- Beyond Eigenfunctions: Divergence Principal Functions for Representation Learning (NeurIPS 2026 · similarity 0.88)
- Learning Neural Eigenfunctions for Unsupervised Semantic Segmentation (ICCV 2023 · similarity 0.87)
- Latent Functional Maps: a spectral framework for representation alignment (NeurIPS 2024 · similarity 0.87)
- Understanding Augmentation-based Self-Supervised Representation Learning via RKHS Approximation and Regression (ICLR 2024 · similarity 0.85)
- Spectral Graph Neural Networks are Incomplete on Graphs with a Simple Spectrum (NeurIPS 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.
