Augmented Cyclic Adversarial Learning for Low Resource Domain Adaptation

Authors: Ehsan Hosseini-Asl, Yingbo Zhou, Caiming Xiong, Richard Socher

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate our proposed model on domain adaptation for visual and speech recognition. We continue the convention of referring to the data domains as source and target , where target denotes the domain with either limited or unlabeled training data. Visual domain adaptation is evaluated using the MNIST dataset (M) Lecun et al. (1998), Street View House Numbers (SVHN) datasets (S) Netzer et al. (2011), USPS (U) (Hull, 1994), MNISTM (MM) and Synthetic Digits (SD) (Ganin and Lempitsky, 2014).
Researcher Affiliation Industry Ehsan Hosseini-Asl, Yingbo Zhou, Caiming Xiong, Richard Socher Salesforce Research {ehosseiniasl,yingbo.zhou,cxiong,rsocher}@salesforce.com
Pseudocode Yes Algorithm 1 Augmented Cyclic Adversarial Learning (ACAL)
Open Source Code No The paper does not provide explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes Visual domain adaptation is evaluated using the MNIST dataset (M) Lecun et al. (1998), Street View House Numbers (SVHN) datasets (S) Netzer et al. (2011), USPS (U) (Hull, 1994), MNISTM (MM) and Synthetic Digits (SD) (Ganin and Lempitsky, 2014). Adaptation on speech is evaluated on the domain of gender within the TIMIT dataset Garofolo et al. (1993), which contains broadband 16k Hz recordings of 6300 utterances (5.4 hours) of phonetically-balanced speech.
Dataset Splits Yes The male/female ratio of speakers across train/validation/test sets is approximately 70% to 30%. Therefore, we treat male speech as the source domain and female speech as the low resource target domain. ... We down sample the MNIST training data so only 10 samples per class are available during training, denoted as MNIST-(10), which is only 0.17% of full training data.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for conducting the experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies, libraries, or frameworks used in the experiments.
Experiment Setup No The paper provides details on model architectures (e.g., 'a modified Le Net consists of two convolutional layers with 20 and 50 channels, followed by a dropout layer and two fully connected layers of 50 and 10 dimensionality') but does not specify key training hyperparameters such as learning rate, batch size, or optimizer settings.