A Convergence Analysis of Gradient Descent on Graph Neural Networks

Authors: Pranjal Awasthi, Abhimanyu Das, Sreenivas Gollapudi

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we verify our theoretical results via simulations.
Researcher Affiliation Industry Pranjal Awasthi Google Research pranjalawasthi@google.com Abhimanyu Das Google Research abhidas@google.com Sreenivas Gollapudi Google Research sgollapu@google.com
Pseudocode No Explanation: The paper does not contain any sections or figures explicitly labeled as "Pseudocode" or "Algorithm" presenting structured steps.
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See supplementary material.
Open Datasets No We generate an unknown ground truth network in the case of Re LU GNNs by choosing each column of W to be a random unit length vector. For the case of linear networks we generate positive definite matrices W 1 , W 2 by picking random Gaussian entries, and then adding a small multiplicative factor of 0.001 times the identity matrix.
Dataset Splits No Explanation: The paper does not specify traditional dataset splits (e.g., training, validation, testing percentages or counts) as it operates in a 'realizable setting' with data generated from an unknown GNN.
Hardware Specification No Our experiments are run using one GPU.
Software Dependencies No We simulate population gradient descent and implement our networks using the JAX programming language [Bradbury et al., 2018].
Experiment Setup Yes For the case of one round GNNs with Re LU activations we set the embedding size r = 10, and h = 10 (number of hidden units in the Re LU GNN). We use the same value of r for the case of deep linear GNNs, where r equals the input dimensionality and also the dimensionality of the matrices W 1 , and W 2 .