Permutation Equivariant Neural Functionals

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

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we find that permutation equivariant neural functionals are effective on a diverse set of tasks that require processing the weights of MLPs and CNNs, such as predicting classifier generalization, producing winning ticket sparsity masks for initializations, and classifying or editing implicit neural representations (INRs).
Researcher Affiliation Academia 1Stanford University 2University of Bologna 3Carnegie Mellon University
Pseudocode Yes A Equivariant NF-Layer pseudocode
Open Source Code Yes In addition, we provide code for our models and experiments1. 1https://github.com/Allan Yang Zhou/nfn
Open Datasets Yes Our first two tasks require (1) predicting the test accuracy of CNN image classifiers and (2) classifying implicit neural representations (INRs) of images and 3D shapes... We consider datasets of SIRENs [58] that encode images (MNIST [38], Fashion MNIST [66], and CIFAR [35]) and 3D shapes (Shape Net-10 and Scan Net-10 [53]).
Dataset Splits Yes Datasets were split into 45,000 training images, 5,000 validation images, and 10,000 (MNIST, CIFAR-10) or 20,000 (Fashion MNIST) test images.
Hardware Specification Yes The model is trained for 50 epochs... which takes 1 hour on a Titan RTX GPU. (Appendix D.1)
Software Dependencies No Appendix A states that the pseudocode uses 'Py Torch [50] and Einops-like [55] pseudocode', but specific version numbers for these software dependencies are not provided.
Experiment Setup Yes Table 12: Hyperparameters for predicting generalization on Small CNN Zoo. Name Values Optimizer Adam Learning rate 0.001 Batch size 8 Loss Binary cross-entropy Epoch 50