Domain Generalization Using a Mixture of Multiple Latent Domains

Authors: Toshihiko Matsuura, Tatsuya Harada11749-11756

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments with benchmark datasets show that our proposed method is effective for domain generalization using a mixture of multiple latent domains, and it outperforms conventional domain generalization methods that use domain labels. Moreover, it is found that the use of pseudo domain labels obtained by clustering style features improves the classification performance compared with the use of original domain labels annotated by humans. To evaluate our proposed method, we perform experiments using two datasets for domain generalization.
Researcher Affiliation Academia Toshihiko Matsuura,1 Tatsuya Harada1,2 1The University of Tokyo, 2RIKEN {matsuura, harada}@mi.t.u-tokyo.ac.jp
Pseudocode Yes Algorithm 1 Training algorithm. Require: Data: Ds = {(xs i, ys i )}Ns i=1 Initialize ˆdi, ˆd i with zero while not end of epoch do Calculate {ddf(xi)}Ns i=1 using Eq. 8 Obtain {ai}Ns i=1 by clustering {ddf(xi)}Ns i=1 Calculate ˆπ using Eq. 9 Update ˆdi with ˆπ(ai) while not end of minibatch do Sample a minibatch of xi, yi, ˆdi Update parameters using Eq. 4 end while Update ˆd i with ˆdi end while
Open Source Code Yes 1The code is publicly available at https://github.com/mil-tokyo/ dg mmld/.
Open Datasets Yes To evaluate our proposed method, we perform experiments using two datasets for domain generalization. PACS (Li et al. 2017a) consists of four domains (i.e., Photo, Art Paintings, Cartoon, and Sketch), spanning different image styles, with seven object categories. VLCS (Torralba and Efros 2011) aggregates images of five shared object categories (bird, car, chair, dog, and person) from PASCAL VOC 2007 (Everingham et al. ), Label Me (Russell et al. 2008), Caltech101 (Fei-Fei, Fergus, and Perona 2007), Sun09 datasets (Choi et al. 2010) which are considered as four separate domains.
Dataset Splits Yes Following the previous work (Carlucci et al. 2019), we use three domains as the source domain, and the other as the target. For the same reason, we split 10% (in the case of PACS) and 30% (in the case of VLCS) of the source samples as validation datasets.
Hardware Specification No The paper does not explicitly describe the specific hardware used for running experiments, such as GPU or CPU models, processor types, or memory details.
Software Dependencies No The paper mentions using pre-trained models (Alex Net, Res Net-18) and a gradient reversal layer, but does not provide specific version numbers for software dependencies such as deep learning frameworks (e.g., PyTorch, TensorFlow), programming languages (e.g., Python), or other libraries.
Experiment Setup Yes Basically, we utilize the other hyper-parameters employed by Ji Gen (Carlucci et al. 2019). In other words, we train the model for 30 epochs using the mini-batch stochastic gradient descent (SGD) with a momentum of 0.9, a weight decay of 5e 4, and a batch size of 128. We set the initial learning rate to 1e 3, and scale it by a factor of 0.1 after 80% of the training epochs. In the experiment with the VLCS dataset, we set the initial learning rate to 1e 4 because it is observed that a high learning rate causes early convergence and overfitting in the source domain. Moreover, we set the learning rate of the classifier and the domain discriminator to be 10 times larger than that of the feature extractor because they are trained from scratch.