AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Permutation Equivariant Neural Functionals

Allan Zhou, Kaien Yang, Kaylee Burns, Adriano Cardace, Yiding Jiang, Samuel Sokota, J Zico Kolter, Chelsea Finn

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
Encode neuron permutation equivariance in the neural functional architecture instead of using statistical weight featuresCNN generalization prediction · Small CNN Zoo: CIFAR-10-GS and SVHN-GSKendall's τ: 0.915 ± 0.002 -> 0.934 ± 0.001; 0.843 ± 0.000 -> 0.931 ± 0.005Improved
Use NFN_NP rather than NFN_HNPINR classification · CIFAR-10, MNIST-10, FashionMNIST, ShapeNet-10, and ScanNet-10test accuracy (%): 44.1 ± 0.471 -> 46.6 ± 0.072; 92.5 ± 0.071 -> 92.9 ± 0.218; 72.7±1.53 -> 75.6±1.07; 86.9±0.860 -> 88.7±0.461; 64.1±0.572Mixed
Use permutation-equivariant NFNs instead of permutation-augmented MLPsINR classification · CIFAR-10, MNIST-10, FashionMNIST, ShapeNet-10, and ScanNet-10test accuracy (%): 18.9 ± 0.432 -> 46.6 ± 0.072; 21.0 ± 0.172 -> 92.9 ± 0.218; 15.9±0.181 -> 75.6±1.07; 39.1 ± 0.385 -> 88.7±0.461; 45.5±0.Improved
Predict winning-ticket masks from initialization with NFN_NPMLP and CNN ticket prediction · MNIST and CIFAR-10test accuracy (%) after training with a 95% sparsity mask: 44.0±0.06 -> 41.4±0.08; 96.2±0.04 -> 94.8±0.01Hurt
Replace NFN_NP's interaction-aware layer with a pointwise layer that ignores weight interactions (ablation)Winning-ticket mask prediction · CIFAR-10 and MNISTtest accuracy (%) after training with a 95% sparsity mask: 41.4±0.08 -> 42.6±0.07; 94.8±0.01 -> 95.0±0.01No clear effect
Remove interactions between weights and layers using a pointwise NF-Layer (ablation)INR weight-space editing · Dilate (MNIST) and Contrast (CIFAR-10)test mean squared errorHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy