A Canonicalization Perspective on Invariant and Equivariant Learning

Authors: George Ma, Yifei Wang, Derek Lim, Stefanie Jegelka, Yisen Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments In this section, we evaluate the expressive power and efficiency of CA on the EXP dataset; we apply our canonicalization to the n-body problem for orthogonal equivariance; we also evaluate OAP for Lap PE on graph regression tasks.
Researcher Affiliation Academia George Ma 1 Yifei Wang 2 Derek Lim2 Stefanie Jegelka3 Yisen Wang4,5 1 School of EECS, Peking University 2 MIT CSAIL 3 TUM CIT/MCML/MDSI & MIT EECS/CSAIL 4 State Key Lab of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University 5 Institute for Artificial Intelligence, Peking University
Pseudocode Yes Algorithm 1 Canonicalization for eliminating sign ambiguity of eigenvectors
Open Source Code Yes Code is available at https://github.com/PKU-ML/canonicalization.
Open Datasets Yes ZINC [24] (MIT License) consists of 12K molecular graphs from the ZINC database of commercially available chemical compounds. EXP [1] (GPL-3.0 License) is a dataset designed to explicitly evaluate the expressiveness of GNN models
Dataset Splits Yes The dataset comes with a predefined 10K/1K/1K train/validation/test split.
Hardware Specification Yes All (preliminary, failed and main) experiments are run on NVIDIA 3090 GPUs with 24GB memory.
Software Dependencies No The Gram-Schmidt process can be implemented in Py Torch [44] using QR decomposition.
Experiment Setup Yes The main hyper-parameters in our experiments are listed as follows. k: the number of eigenvectors used in the PE. L1: the number of layers of the base model. h1: the hidden dimension of the base model. h2: the output dimension of the base model. λ: the initial learning rate. t: the patience of the learning rate scheduler. r: the factor of the learning rate scheduler. λmin: the minimum learning rate of the learning rate scheduler. L2: the number of layers of Sign Net or the normal GNN6 (when using canonicalization as PE). h3: the hidden dimension of Sign Net or the normal GNN (when using canonicalization as PE). The values of these hyper-parameters in our experiments are listed in Table 9.