Learning Invariances in Neural Networks from Training Data

Authors: Gregory Benton, Marc Finzi, Pavel Izmailov, Andrew G. Wilson

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For the experiments in this and all following sections we use a 13-layer CNN architecture from Laine and Aila [21].", "we achieve 98.9% test accuracy.", "Table 1: Test accuracy for models trained on CIFAR-10 with different augmentations applied to the training data.", "Table 2: Test MAE (in meV) on QM9 tasks trained with specified augmentation.
Researcher Affiliation Academia Gregory Benton Marc Finzi Pavel Izmailov Andrew Gordon Wilson Courant Institute of Mathematical Sciences New York University
Pseudocode Yes Algorithm 1: Learning Invariances with Augerino
Open Source Code Yes The accompanying code can be found at https://github.com/g-benton/learning-invariances.
Open Datasets Yes The rotated MNIST dataset (rot MNIST)", "Using the CIFAR-10 dataset", "We test out our method on the molecular property prediction dataset QM9 [3, 34]", "construct rot Cam Vid, a variation of the Cam Vid dataset [5, 4]", "learn a uniform distribution over the brightness and contrast adjustments on STL-10 [6].
Dataset Splits No The paper mentions 'training set' and 'test set' for some datasets but does not explicitly provide specific train/validation/test percentages or counts needed for reproduction in the main text.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments are provided in the paper.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., library names with their versions) that are needed to replicate the experiment.
Experiment Setup Yes For the experiments in this and all following sections we use a 13-layer CNN architecture from Laine and Aila [21].", "We compare Augerino trained with three values of λ from Equation 5; λ = {0.01, 0.05, 0.1}", "training for 200 epochs", "We train the models for 500 epochs on MAE.