If You Want to Be Robust, Be Wary of Initialization

Authors: Sofiane ENNADIR, Johannes Lutzeyer, Michalis Vazirgiannis, El Houcine Bergou

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments, spanning diverse models and real-world datasets subjected to various adversarial attacks, validate our findings.
Researcher Affiliation Academia Sofiane Ennadir KTH Stockholm, Sweden Johannes F. Lutzeyer LIX, Ecole Polytechnique IP Paris, France Michalis Vazirgiannis KTH & Ecole Polytechnique Stockholm, Sweden El Houcine Bergou UM6P Benguerir, Morocco
Pseudocode No The paper describes methods through mathematical formulations and textual descriptions of processes, but it does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes The necessary code to reproduce all our experiments is available on github https://github.com/Sennadir/Initialization_effect.
Open Datasets Yes We leverage the citation networks Cora and Cite Seer [27], with additional results on other datasets provided in the Appendix G.
Dataset Splits Yes To mitigate the impact of randomness during training, each experiment was repeated 10 times, using the train/validation/test splits provided with the datasets.
Hardware Specification Yes The experiments have been run on both a NVIDIA A100 GPU where training a GCN takes around 1.2( 0.2) s.
Software Dependencies No Our implementation is built using the open-source library Py Torch Geometric (Py G) under the MIT license [12].
Experiment Setup Yes We maintained the same hyperparameters, including a learning rate of 1e-2, 300 epochs, and a hidden feature dimension of 16 have been.