Transferable Curriculum for Weakly-Supervised Domain Adaptation

Authors: Yang Shu, Zhangjie Cao, Mingsheng Long, Jianmin Wang4951-4958

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments We evaluate TCL with state-of-the-art curriculum schemes and deep domain adaptation methods on three datasets.
Researcher Affiliation Academia Yang Shu, Zhangjie Cao, Mingsheng Long,B Jianmin Wang School of Software, Tsinghua University, China KLiss, MOE; BNRist; Research Center for Big Data, Tsinghua University, China {shuyang5656,caozhangjie14}@gmail.com {mingsheng,jimwang}@tsinghua.edu.cn
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes Code and datasets will be available at github.com/thuml.
Open Datasets Yes Office-31 (Saenko et al. 2010) is a standard dataset for domain adaptation... Office-Home (Venkateswara et al. 2017) is a more challenging dataset for visual domain adaptation... Bing-Caltech (Bergamo and Torresani 2010) was created with Bing and Caltech-256 datasets.
Dataset Splits No The paper mentions selecting hyperparameters using 'cross validation' but does not provide specific dataset split percentages, sample counts, or a detailed splitting methodology for training, validation, and testing sets.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions 'Py Torch' as the implementation framework but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We use Res Net-50 pre-trained on the Image Net dataset (Russakovsky et al. 2015) as our base model, and add a fullyconnected bottleneck layer before its classifier layer. Since the dataset is relatively small and the source domain is noisy, we fine-tune only the last residual block of the Res Net-50 model, and train the bottleneck layer, the classifier layer and the domain discriminator from scratch. Before using the curriculum, we pre-train our network on noisy data for a few epochs, which is better than random initialization. The tradeoff hyper-parameter λ is selected according to magnitudes of the two terms in Eq. (6), and the threshold γ is selected according to the distribution of loss values using cross validation. We use mini-batch SGD with momentum of 0.9 and the same learning rate strategy in (Ganin et al. 2016).