Approximately Equivariant Graph Networks
Authors: Ningyuan Huang, Ron Levie, Soledad Villar
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate our theory in three real-world tasks for learning on a fixed graph: image inpainting, traffic flow prediction, and human pose estimation. We show theoretically and empirically that the best generalization performance can be achieved by choosing a suitably larger group than the graph automorphism, but smaller than the permutation group. |
| Researcher Affiliation | Academia | Ningyuan (Teresa) Huang Johns Hopkins University nhuang19@jhu.edu Ron Levie Technion Israel Institute of Technology levieron@technion.ac.il Soledad Villar Johns Hopkins University svillar3@jhu.edu |
| Pseudocode | Yes | Algorithm 1 Parameterizing equivariant linear functions f : RN RN for abelian group |
| Open Source Code | Yes | The source code is available at https://github.com/nhuang37/Approx_Equivariant_Graph_Nets. |
| Open Datasets | Yes | We use subsets of MNIST [74] and Fashion MNIST [75], each comprising 100 training samples and 1000 test samples. The traffic graph is typically large and asymmetric. Therefore we leverage our approximate symmetries to study the symmetry model selection problem (Section 4). We use the METR-LA dataset which represents traffic volume of highways in Los Angeles (see figure inset). We use the standard benchmark dataset Human3.6M [77] and follow the evaluation protocol in [78]. |
| Dataset Splits | Yes | We report the best test accuracy at the model checkpoint selected by the best validation accuracy (with a 80/20 training-validation split). We use the same traffic data normalization and 70/10/20 train/validation/test data split as [76]. |
| Hardware Specification | Yes | All experiments were conducted on a server with 256 GB RAM and 4 NVIDIA RTX A5000 GPU cards. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | We train the models with ADAM (learning rate 0.01, no weight decay, at most 1000 epochs). We train all variants for 30 epochs using ADAM optimizer with learning rate 0.01. We design G-Net to have 4 layers (with batch normalization and residual connections in between the hidden layers), 128 hidden units, and use ReLU nonlinearity. We train our models for at most 30 epochs with early stopping. For comparison purpose, we use the same optimization routines as in Sem GCN [78] and perform the hyper-parameter search of learning rates {0.001, 0.002}. |