Deep invariant networks with differentiable augmentation layers

Authors: Cédric ROMMEL, Thomas Moreau, Alexandre Gramfort

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide empirical evidence showing that our approach is easier and faster to train than modern automatic data augmentation techniques based on bilevel optimization, while achieving comparable results.
Researcher Affiliation Academia Cédric Rommel, Thomas Moreau & Alexandre Gramfort Université Paris-Saclay, Inria, CEA, Palaiseau, 91120, France {firstname.lastname}@inria.fr
Pseudocode No The paper describes the architecture and method but does not include a formal pseudocode or algorithm block.
Open Source Code Yes The accompanying code can be found at https://github.com/cedricrommel/augnet.
Open Datasets Yes In this experiment we showcase Aug Net on a standard image recognition task using the CIFAR10 dataset [27].
Dataset Splits Yes All models considered in this experiment are trained for 300 epochs over 5 different seeds on a random 80% fraction of the official CIFAR10 training set. The remaining 20% is used as a validation set for early-stopping and choosing hyperparameters, and the official test set is used for reporting performance metrics.
Hardware Specification Yes It was also granted access to the HPC resources of IDRIS under the allocation 2021-AD011012284R1 and 2022-AD011011172R2 made by GENCI.
Software Dependencies No Our work is based on code from [19] and [12], as well as open-source libraries like MNE-PYTHON [35] and BRAINDECODE [36].
Experiment Setup Yes All models considered in this experiment are trained for 300 epochs over 5 different seeds on a random 80% fraction of the official CIFAR10 training set. The reader is referred to Section A.3 for further experimental details, and to Section B.2 for a sensitivity analysis on hyperparameters C and λ.