Research map / Architectures and efficiency / RNNs and normalization
Neural Functional Transformers
Allan Zhou, Kaien Yang, Yiding Jiang, Kaylee Burns, Winnie Xu, Samuel Sokota, J Zico Kolter, Chelsea Finn
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 NFN weight-space layers with attention-based NFT layers | INR editing · MNIST erosion, dilation, and gradient; FashionMNIST gradient; CIFAR contrast | test MSE: 0.0223 ± 0.0000 -> 0.0194 ± 0.0002; 0.0693 ± 0.0009 -> 0.0510 ± 0.0004; 0.0566±0.0000 -> 0.0484±0.0007; 0.0870 ± 0.0001 | Improved |
| use NFT rather than NFN_HNP for CNN generalization prediction | CNN generalization predictor · CIFAR-10-GS | Kendall rank correlation τ: 0.934 ± 0.001 -> 0.926 ± 0.001 | Hurt |
| use NFT rather than NFN_HNP for CNN generalization prediction | CNN generalization predictor · SVHN-GS | Kendall rank correlation τ: 0.931 ± 0.005 -> 0.858 ± 0.000 | Hurt |
| replace the NFN encoder with an NFT encoder (ablation) | Inr2Array · MNIST INR classification | test accuracy (%): 94.6 ± 0.00 -> 98.5 ± 0.00 | Improved |
| replace the NFN encoder with an NFT encoder (ablation) | Inr2Array · FashionMNIST INR classification | test accuracy (%): 76.7 ± 0.00 -> 79.3 ± 0.00 | Improved |
| replace the NFN encoder with an NFT encoder (ablation) | Inr2Array · CIFAR-10 INR classification | test accuracy (%): 45.4 ± 0.00 -> 63.4 ± 0.00 | Improved |
| add spatial activations without Inr2Array reconstruction training (ablation) | NFN INR classifier · MNIST, FashionMNIST, and CIFAR-10 INR classification | test accuracy (%): 92.9±0.46 vs 92.9 ± 0.38; 70.8±0.53 vs 75.6 ± 1.07; 45.6 ± 0.11 vs 46.6 ± 0.13 | Hurt |
| replace the Inr2Array NFT encoder with an NFN encoder (ablation) | Inr2Array reconstruction · MNIST, FashionMNIST, and CIFAR-10 | test MSE: 0.027 ± 0.003 -> 0.046 ± 0.001; 0.070 ± 0.006 -> 0.085 ± 0.006; 0.036 ± 0.006 -> 0.117 ± 0.008 | Hurt |
Closest papers by meaning
- Universal Neural Functionals (NeurIPS 2024 · similarity 0.91)
- Equivariant Neural Functional Networks for Transformers (ICLR 2025 · similarity 0.89)
- Few-shot Implicit Function Generation via Equivariance (CVPR 2025 · similarity 0.87)
- Equivariant Polynomial Functional Networks (ICML 2025 · similarity 0.87)
- Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion (ICML 2025 · similarity 0.86)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.86)
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.
