Task Cooperation for Semi-Supervised Few-Shot Learning

Authors: Han-Jia Ye, Xin-Chun Li, De-Chuan Zhan10682-10690

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The state-of-the-art few-shot classification results on Mini Image Net and Tiered Image Net verify the superiority of TACO to leverage unlabeled data and task relationship in meta-learning.
Researcher Affiliation Academia Han-Jia Ye, Xin-Chun Li, De-Chuan Zhan State Key Laboratory for Novel Software Technology, Nanjing University {yehj, lixc, zhandc}@lamda.nju.edu.cn
Pseudocode Yes Algorithm 1 The meta-training flow of the TACO.
Open Source Code No The paper does not provide an explicit statement or link indicating that the source code for the methodology is openly available.
Open Datasets Yes Datasets. Mini Image Net (Vinyals et al. 2016) and Tiered Image Net (Ren et al. 2018) contain 100 classes and 608 classes respectively.
Dataset Splits Yes The first strategy splits all examples in the meta-train set across classes (SAC). In this case, we randomly select 30% classes in the meta-train set as the labeled part and uses the instances in the remaining classes without their labels as the unlabeled set. Similarly, we randomly select 30% instances across instances (SAI). In the SAI case, it is possible to sample non-distractor classes from the unlabeled pool, which reduces the classification difficulty w.r.t. SAC to some extent. ... Thus instead of preserving the whole meta-val set, we adopt the same SAC or SAI split methods to reduce the size of the meta-val set.
Hardware Specification No The paper does not specify any particular hardware components such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper mentions '4-layer Conv Net' and 'Res Net-12' as backbones and 'Proto Net' for implementation, but does not provide specific version numbers for any software dependencies like programming languages, libraries, or frameworks.
Experiment Setup Yes Implementation Details. We use a 4-layer Conv Net (Vinyals et al. 2016; Finn, Abbeel, and Levine 2017; Snell, Swersky, and Zemel 2017) as the backbone... For semi-supervised FSL, we sample 75 unlabeled instances in each mini-batch.