Understanding Cross-Domain Few-Shot Learning Based on Domain Similarity and Few-Shot Difficulty

Authors: Jaehoon Oh, Sungnyun Kim, Namgyu Ho, Jin-Hwa Kim, Hwanjun Song, Se-Young Yun

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we empirically investigate which pre-training is preferred based on domain similarity and few-shot difficulty of the target domain. ... Our code is available at https://github.com/sungnyun/understanding-cdfsl. ... We conduct an empirical study to gain an in-depth understanding of their effectiveness in the pre-training phase...
Researcher Affiliation Collaboration KAIST DS Daejeon, South Korea ... KAIST AI Seoul, South Korea ... NAVER AI Lab, SNU AIIS ... NAVER AI Lab
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Our code is available at https://github.com/sungnyun/understanding-cdfsl.
Open Datasets Yes We use Image Net, tiered Image Net, and mini Image Net as source datasets for generality. Regarding the target domain, we prepare eight datasets... BSCDFSL (Broader Study of CD-FSL) benchmark [24]... Euro SAT [27]... Image Net [11]... CUB [70]... Cars [34]... Plantae [39]... ISIC [7]... Chest X [68]... Places [74].
Dataset Splits Yes We follow the split strategy used in Phoo and Hariharan [47], where 20% of the target data DN is used as the unlabeled data DU for pre-training. ... We use 20% of the target dataset as labeled data to pre-train the model in a supervised manner. Then, the pre-trained model is evaluated on the remaining unseen target data for the 5-way k-shot classification task. ... The support set DS and query set DQ consist of n classes that are randomly selected among the entire set of novel classes CN. For the n-way k-shot setting, k examples are randomly drawn from each class for the support set DS, while kq (typically 15) examples for the query set DQ.
Hardware Specification Yes All models are trained on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions 'Py Torch [45]' but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes The balancing coefficient γ in Eq. (3) of MSL is set to be 0.875 found by a grid search... We use Adam optimizer [33] with β1 = 0.9 and β2 = 0.999. The learning rate is initialized as 10 3 and decreases by 0.1 at 80th and 90th epochs. ... We use different backbone networks depending on the source data. For Image Net and tiered Image Net, Res Net18 is used as the backbone, while Res Net10 is used for mini Image Net.