Learning to Self-Train for Semi-Supervised Few-Shot Classification

Authors: Xinzhe Li, Qianru Sun, Yaoyao Liu, Qin Zhou, Shibao Zheng, Tat-Seng Chua, Bernt Schiele

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed LST method in terms of few-shot image classification accuracy in semisupervised settings. Below we describe the two benchmarks we evaluate on, details of settings, comparisons to state-of-the-art methods, and an ablation study.
Researcher Affiliation Collaboration Xinzhe Li1 Qianru Sun2 Yaoyao Liu3 Shibao Zheng1 Qin Zhou4 Tat-Seng Chua5 Bernt Schiele6 1Shanghai Jiao Tong University 2Singapore Management University 3Tianjin University 4Alibaba Group 5National University of Singapore 6Max Planck Institute for Informatics, Saarland Informatics Campus
Pseudocode No The paper includes diagrams illustrating the method's pipeline and training procedures (Figure 1, Figure 2), but these are not structured pseudocode or algorithm blocks.
Open Source Code Yes Code is at github.com/xinzheli1217/learning-to-self-train.
Open Datasets Yes We conduct our experiments on two subsets of Image Net [26]. mini Image Net was firstly proposed by Vinyals et al. [36] ... tiered Image Net was proposed by Ren et al. [24].
Dataset Splits Yes In the uniform setting, these classes are divided into 64, 16, and 20 respectively for meta-train, meta-validation, and meta-test.
Hardware Specification No The paper describes network architectures (e.g., Res Net-12) and general computational processes but does not provide specific hardware details (e.g., exact GPU/CPU models) used for running its experiments.
Software Dependencies No The paper mentions common deep learning frameworks and libraries implicitly, but it does not specify any software components with their version numbers required for reproducibility.
Experiment Setup Yes Base-learning rate α (in Eq. 1, Eq. 4 and Eq. 6) is set to 0.01. Meta-learning rates β1 and β2 (in Eq. 7 and Eq. 8) are set to 0.001 initially and decay to the half value every 1k meta iterations until a minimum value 0.0001 is reached. We use a meta-batch size of 2 and run 15k meta iterations. In recursive training, we use 6 (3) recursive stages for 1-shot (5-shot) tasks. Each recursive stage contains 10 re-training and 30 fine-tuning steps.