Task-Level Self-Supervision for Cross-Domain Few-Shot Learning
Authors: Wang Yuan, Zhizhong Zhang, Cong Wang, Haichuan Song, Yuan Xie, Lizhuang Ma3215-3223
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the effectiveness of the proposed T3S under two cross-domain few-shot settings. First, we train the few-shot classification model on the mini Image Net and test the trained model on other four different benchmarks. Second, we design a generalized task-agnostic test, where we re-think the generalization ability of existing FSL methods. and Table 1: Classification accuracy on four cross domain experiments. |
| Researcher Affiliation | Collaboration | Wang Yuan1 , Zhizhong Zhang1 , Cong Wang3, Haichuan Song1, Yuan Xie1*, Lizhuang Ma1,2,4* 1 East China Normal University 2 Shanghai Jiao Tong University 3 Huawei Technologies 4 Mo E Key Lab of Artificial Intelligence, Shanghai Jiao Tong University |
| Pseudocode | No | The paper describes the methodology in text and with diagrams but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link for the open-source code of the proposed T3S method. It only mentions using official code or re-implementing competitors' code. |
| Open Datasets | Yes | Source Dataset: mini Image Net (Vinyals et al. 2016), a subset of the ILSVRC-12 (Deng et al. 2009), is a standard benchmark for few-shot image classification. It consists of 60,000 color images of size 84x84 with 100 classes. |
| Dataset Splits | Yes | We follow the splitting introduced by (Ravi and Larochelle 2017), with 64, 16, and 20 classes for training, validation and testing, respectively. We take the training set as source domain and select the model on the validation data. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'C3D' but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | We totally train 1,000K episodes for our model using the Adam optimizer with initial learning rate 10^-3 and exponentially decayed by 50% every 50k episode. In each episode, we sample N K support samples and 16 query samples. The mini-batch size is empirically set to be 64 for 5-way 1-shot/5-shot task. |