Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images

Authors: Thomas Lucas, Philippe Weinzaepfel, Gregory Rogez1881-1889

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that our approach performs significantly better on STL-10 in the barely-supervised regime, e.g. with 4 or 8 labeled images per class. ... Summary of our main contributions: An analysis of the distillation dilemma in Fix Match. ... Experiments showing that our approach allows barelysupervised learning on the more realistic STL-10 dataset.
Researcher Affiliation Industry Thomas Lucas1, Philippe Weinzaepfel1, Gregory Rogez1 1Naver Labs Europe*
Pseudocode No The paper includes a high-level overview diagram in Figure 1, but no detailed pseudocode or algorithm blocks are present.
Open Source Code No The paper does not contain any explicit statements or links indicating that source code for the described methodology is publicly available.
Open Datasets Yes We perform most ablations on STL-10 and also compare approaches on CIFAR-10 and CIFAR-100. ... The STL-10 dataset consists of 5k labeled images of resolution 96 96 split into 10 classes, and 100k unlabeled images.
Dataset Splits Yes We use various amounts of labeled data: 10 (1 image per class), 20, 40, 80, 250, 1000. ... We average across 4 random seeds for 4 images per class or less, 3 otherwise, and across the last 10 checkpoints of all runs.
Hardware Specification No The paper mentions the use of Wide-Res Net architectures (WR-28-2, WR-28-8, WR-37-2) but does not specify any hardware details like GPU models, CPU types, or other computing resources used for experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies (e.g., programming languages, libraries, frameworks) used in the experiments.
Experiment Setup Yes We use τ = 0.95 for Fix Match and τ = 0.98 for our model, see Section 5.3 for discussions about setting τ. ... Standard deviations increase as the number of labels decreases, so we average across 4 random seeds for 4 images per class or less, 3 otherwise, and across the last 10 checkpoints of all runs.