Research map / Learning theory and algorithms / Kernels and regression theory
NeuralEF: Deconstructing Kernels by Deep Neural Networks
Zhijie Deng, Jiaxin Shi, 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 |
|---|---|---|---|
| Approximate the CNN-GP kernel using learned top eigenfunctions | MNIST image projections with logistic regression · MNIST test | LR test accuracy: 78.00% -> 84.98% (polynomial kernel) | Improved |
| Use NeuralEF-learned NTK projections instead of random features | K-means clustering · CIFAR-10 test images | accuracy: 81.3% -> 97.5% | Improved |
| Use NeuralEF-learned NTK projections instead of raw pixels | K-means clustering · CIFAR-10 test images | accuracy: 67.8% -> 97.5% | Improved |
| Use NeuralEF to approximate the NTK in linearised Laplace approximation | CIFAR-10 ResNet-20 LLA · CIFAR-10 test | ECE: 0.049 -> 0.016 (MAP); 0.026 -> 0.016 (Last-layer LLA) | Improved |
| Use NeuralEF to approximate the NTK in linearised Laplace approximation | CIFAR-10 ResNet-56 LLA · CIFAR-10 test | ECE: 0.024 -> 0.012 (Last-layer LLA) | Improved |
| Use NeuralEF to approximate the NTK in linearised Laplace approximation | CIFAR-10 ResNet-110 LLA · CIFAR-10 test | ECE: 0.019 -> 0.010 (Last-layer LLA) | Improved |
| Use NeuralEF LLA rather than Last-layer LLA | CIFAR-10 ResNet-20 · CIFAR-10 test | NLL: 0.264 -> 0.277 | Hurt |
| Use NeuralEF LLA rather than Last-layer LLA | CIFAR-10 ResNet-56 · CIFAR-10 test | NLL: 0.231 -> 0.234 | Hurt |
| Use NeuralEF LLA rather than Last-layer LLA | CIFAR-10 ResNet-110 · CIFAR-10 test | NLL: 0.233 -> 0.241 | Hurt |
| Approximate the NTK using only the top-10 eigenpairs | NeuralEF kernel recovery · ResNet-20 binary-classifier NTK test kernel | kernel approximation quality | Hurt |
| Use NeuralEF instead of random features at similar resource consumption | NTK kernel approximation · ResNet-20 binary-classifier NTK | approximation quality | Improved |
Later papers that compared with this method
1 later accepted paper reports NeuralEF as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- STNet: Spectral Transformation Network for Solving Operator Eigenvalue Problem (NeurIPS 2025 · 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
- Neural Eigenfunctions are Structured Representation Learners (TPAMI 2025 · similarity 0.89)
- Improved Operator Learning by Orthogonal Attention (ICML 2024 · similarity 0.86)
- On the Eigenvalue Decay Rates of a Class of Neural-Network Related Kernel Functions Defined on General Domains (JMLR 2024 · similarity 0.86)
- Eigen Analysis of Conjugate Kernel and Neural Tangent Kernel (ICML 2025 · similarity 0.86)
- Improving Adaptivity via Over-Parameterization in Sequence Models (NeurIPS 2024 · similarity 0.85)
- Shortcut Features as Top Eigenfunctions of NTK: A Linear Neural Network Case and More (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.
