Adversarial Dropout for Recurrent Neural Networks

Authors: Sungrae Park, Kyungwoo Song, Mingi Ji, Wonsung Lee, Il-Chul Moon4699-4706

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental According to our experiments, adversarial dropout for RNNs showed the advanced performances on the sequential versions of MNIST, the semi-supervised text classification, and the language modeling tasks. ... Table 1 shows the test performances of the dropout-based regularizations.
Researcher Affiliation Collaboration Sungrae Park,1 Kyungwoo Song,2 Mingi Ji,2 Wonsung Lee,3 Il-Chul Moon2 1Clova AI Research, NAVER Corp., Korea 2Industrial & Systems Engineering, KAIST, Korea 3AI Center, SK Telecom, Korea
Pseudocode No The paper describes algorithmic steps and refers to a 'detail algorithm in appendix' but does not present a clearly labeled pseudocode or algorithm block within the provided text.
Open Source Code Yes Our implementation code will be available at https://github.com/sungraepark/adversarial dropout text classification. ... Our implementation code will be available at https://github.com/sungraepark/adversarial dropout lm.
Open Datasets Yes Sequential MNIST tasks, also known as pixel-by-pixel MNIST... IMDB is a standard benchmark movie review dataset... Elec is a dataset on electronic product reviews from Amazon... the Penn Treebank (PTB)... and Wiki Text-2 (WT2) dataset...
Dataset Splits Yes These hyperparameters of the baseline models as well as our models were retrieved in the validation phase. ... Table 3 shows the perplexity on both the PTB and Wiki Text-2 validation and test datasets.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU model, GPU model, memory size) used to conduct the experiments.
Software Dependencies No The paper mentions implementing models but does not list specific software dependencies (e.g., libraries, frameworks) along with their version numbers.
Experiment Setup Yes For the settings for the dropout, we set the dropout probability as 0.1 for the baseline models. In the case of the adversarial dropout, we adapted ϵ0 = Eϵ[ϵ], which indicates the expectation of the dropout mask, and δ = 0.03, which represents the maximum changes from the base dropout mask as 3%. These hyperparameters of the baseline models as well as our models were retrieved in the validation phase. All models were trained with the same optimizer (detail settings in appendix).