KL Guided Domain Adaptation
Authors: A. Tuan Nguyen, Toan Tran, Yarin Gal, Philip Torr, Atilim Gunes Baydin
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results also suggest that our method outperforms other representation-alignment approaches.We conduct extensive experiments and show that our method significantly outperforms relevant baselines, namely ERM (Bousquet et al., 2003), DANN (Ganin et al., 2016), MMD (Gretton et al., 2012; Li et al., 2018), CORAL (Sun and Saenko, 2016) and WD (Shen et al., 2018). |
| Researcher Affiliation | Collaboration | University of Oxford, Oxford, United Kingdom Vin AI Research, Hanoi, Vietnam |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | For the full description of these baselines, please refer to our appendix and the official code at https://github.com/atuannguyen/KL. |
| Open Datasets | Yes | Rotated MNIST consists of 70,000 MNIST (Le Cun et al., 2010) images...DIGITS is a common domain adaptation dataset, with 3 digit classification sub-datasets, namely MNIST, USPS (Hull, 1994) and SVHN (Netzer et al., 2011).Vis DA17 (Peng et al., 2017) is a challenging real-world classification dataset |
| Dataset Splits | Yes | We use the remaining 20% of the source data as the validation set, and the remaining 20% of the target domain data as the test set. |
| Hardware Specification | Yes | We train all models on an NVIDIA Quadro RTX 6000 GPU. |
| Software Dependencies | No | The paper mentions 'Pytorch (Paszke et al., 2019)' but does not provide a specific version number for PyTorch or any other software dependencies. It also mentions 'Adam optimizer Kingma and Ba (2014)' but this is an algorithm, not a software dependency with a version. |
| Experiment Setup | Yes | For the Rotated MNIST and DIGITS experiments, we use a simple convolutional neural network with four 3 3 convolutional layers (followed by an average pooling layer) as the representation network.We train each model for 100 epochs.tune the hyperparameters (learning rate, regularizer coefficients, weight decay, representation dimension and dropout rate)learning rate: [10 4.5, 10 2.5], weight decay: 0.0, dropout rate: 0.0, batch size: [8, 512] |