Rectifying the Shortcut Learning of Background for Few-Shot Learning
Authors: Xu Luo, Longhui Wei, Liangjian Wen, Jinrong Yang, Lingxi Xie, Zenglin Xu, Qi Tian
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we for the first time empirically identify image background, common in realistic images, as a shortcut knowledge helpful for in-class classification but ungeneralizable beyond training categories in FSL. Extensive experiments carried on inductive FSL tasks demonstrate the effectiveness of our approaches. |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China 2University of Science and Technology of China 3Tsinghua University 4Huazhong University of Science and Technology 5Xidian University 6Harbin Institute of Technology Shenzhen 7Pengcheng Laboratory |
| Pseudocode | No | While the paper describes algorithms (COS, SOC) in a step-by-step manner, these descriptions are not formatted as explicit pseudocode blocks or labeled as 'Algorithm'. |
| Open Source Code | Yes | Code: https://github.com/Frankluox/Few Shot Code Base |
| Open Datasets | Yes | The first is mini Image Net [53], a small subset of ILSVRC-12 [44] that contains 600 images within each of the 100 categories. The second dataset, tiered Image Net [41], is a much larger subset of ILSVRC-12 |
| Dataset Splits | Yes | The categories are split into 64, 16, 20 classes for training, validation and evaluation, respectively. The super-classes are split into 20, 6, 8 super-classes which ensures separation between training and evaluation categories. |
| Hardware Specification | Yes | We use Pytorch [38] to implement all our experiments on two NVIDIA 1080Ti GPUs. |
| Software Dependencies | No | The paper mentions 'Pytorch [38]' but does not specify a version number for this or any other software dependency. |
| Experiment Setup | Yes | The initial learning rate for training Exemplar is 0.1, and for CC is 0.005. The batch size for Exemplar, CC are 256 and 128, respectively. For mini Image Net, we train Exemplar for 150k iterations... The threshold γ is set to 0.5, and top 3 out of 30 features are chosen per image at the training stage. At the evaluation stage, we crop each image 7 times. The importance factors α and β are both set to 0.8. |