Understanding Noise-Augmented Training for Randomized Smoothing
Authors: Ambar Pal, Jeremias Sulam
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our analysis has direct implications to the practical deployment of randomized smoothing, and we illustrate some of these via experiments on CIFAR-10 and MNIST, as well as on synthetic datasets. |
| Researcher Affiliation | Academia | Ambar Pal EMAIL Department of Computer Science & Mathematical Institute of Data Science Johns Hopkins University Baltimore, MD 21218, USA |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide code for all of our experiments at https://github.com/ambarpal/randomized-smoothing. |
| Open Datasets | Yes | Our empirical experiments suggest, firstly, that real data distributions lie in the low-interference distance regime, and hence noise-augmented training is beneficial to randomized smoothing. Secondly, contrary to practice, our proofs suggest that the parameters of the noise-distribution for noiseaugmented training and that of randomized smoothing need not be the same, and that allowing for different values of these parameters lead to improved results. Our experiments on MNIST and CIFAR-10 confirm this theoretical intuition. |
| Dataset Splits | Yes | We take 100 samples from each of pα, pβ per image x in the MNIST test set to compute a single point α,β. The figures are quite stable against random initialization due to the MNIST test set size of 10000 samples. We report results aggregated over 4000 images from the CIFAR-10 test set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment. |
| Experiment Setup | Yes | For a range of finely spaced α in [0, 1], we train a standard CNN hα with isotropic Gaussian noise-augmentation with variance α2, i.e., pα = N(0, α2I). For each of these trained models, we use the isotropic-Gaussian with variance β2, with β [0, 1], as the smoothing distribution, i.e., pβ = N(0, β2I), to obtain the smoothed classifier Smoothβ(ψ(h pα)). We use the following architecture for the neural network classifier hα: Input(28, 28) Conv3,1(1, 32) Conv3,1(32, 64) Linear(9216, 128) Linear(128, 2). |