Hierarchical Randomized Smoothing

Authors: Yan Scholten, Jan Schuchardt, Aleksandar Bojchevski, Stephan Günnemann

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally demonstrate the importance of hierarchical smoothing in image and node classification, where it yields superior robustness-accuracy trade-offs.
Researcher Affiliation Academia 1Dept. of Computer Science & Munich Data Science Institute, Technical University of Munich 2University of Cologne, Germany
Pseudocode Yes Algorithm 1: Binary-class Robustness Certificates for Hierarchical Randomized Smoothing Algorithm 2: Multi-class Robustness Certificates for Hierarchical Randomized Smoothing
Open Source Code Yes Project page: https://www.cs.cit.tum.de/daml/hierarchical-smoothing To ensure reproducibility we further use random seeds for all random processes including for model training, for drawing smoothing parameters and for drawing Monte-Carlo samples. We publish the source code including reproducibility instructions and all random seeds.
Open Datasets Yes For image classification we train Res Net50 (He et al., 2016) on CIFAR10 (Krizhevsky et al., 2009)... For node classification we train graph attention networks (GATs) (Velickovic et al., 2018) with two layers on Cora-ML (Mc Callum et al., 2000; Bojchevski and Günnemann, 2018)... The graph datasets are publicly available and can be downloaded also e.g. using Py Torch Geometric. The CIFAR10 dataset is also publicly available, for example via the torchvision library.
Dataset Splits Yes We label 20 randomly drawn nodes per class for training and validation, and 10% of the nodes for testing. We train our models on training nodes, tune them on validation nodes and compute certificates for all test nodes.
Hardware Specification Yes All experiments were performed on a Xeon E5-2630 v4 CPU with a NVIDIA GTX 1080TI GPU.
Software Dependencies No The paper mentions specific libraries like "Py Torch Geometric" and "torchvision library" and algorithms like "Adam", but does not provide specific version numbers for all key software components (e.g., PyTorch, Python versions).
Experiment Setup Yes At test time, we use significance level α = 0.01, n0 = 1,000 samples for estimating the majority class and n1 = 10,000 samples for certification. ... We train models full-batch using Adam (learning rate = 0.001, β1 =0.9, β2 =0.999, ϵ = e 08, weight decay = 5e 04) for a maximum of 1,000 epochs with cross-entropy loss. ... We train the Res Net50 model with stochastic gradient descent (learning rate 0.01, momentum 0.9, weight decay 5e-4) for 400 epochs using a batch-size of 128.