Meta Variance Transfer: Learning to Augment from the Others

Authors: Seong-Jin Park, Seungju Han, Ji-Won Baek, Insoo Kim, Juhwan Song, Hae Beom Lee, Jae-Joon Han, Sung Ju Hwang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our model on multiple benchmark datasets for few-shot classification and face recognition, on which our model significantly improves the performance of the base model, outperforming relevant baselines.
Researcher Affiliation Collaboration 1Samsung Advanced Institute of Technology 2KAIST.
Pseudocode Yes Algorithm 1 Meta Variance Transfer
Open Source Code No The paper does not contain an explicit statement about providing source code for the described methodology or a link to a public repository.
Open Datasets Yes We extensively validate our method on diverse few-shot classification datasets. For fine-grained few-shot classification, we use CUB dataset (Wah et al., 2011) and face recognition datasets: VGGface2 (Cao et al., 2018) and CASIA-webface (Yi et al., 2014) datasets. We also use mini Image Net dataset (Deng et al., 2009; Vinyals et al., 2016; Ravi & Larochelle, 2016) which is widely used as a few-shot meta-learning benchmark.
Dataset Splits Yes We randomly select 300 subjects for each dataset and split them to 200, 50, and 50 subjects for seen, validation, and unseen datasets, respectively.
Hardware Specification Yes We used a single NVIDIA P40 GPU for training the model.
Software Dependencies No The paper mentions 'Py Torch implementation' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We use 84 84 images for the Conv4 network, and 224 224 images for Res Net-based networks... We train the model by using the Adam optimizer (Kingma & Ba, 2014) for about 110k iterations with the initial learning rate of 0.001, and decay it by 0.1 at 70k and 90k iterations. We set the number of queries during both meta-training and metatesting as 15... For the hyper-parameter λ in equation 1, we used 0.1 throughout the experiments.