A Survey on Semi-, Self- and Unsupervised Learning for Image Classification

TL;DR

This survey reviews semi-, self-, and unsupervised learning for image classification, comparing 34 methods by their shared ideas and reported performance rather than a fine-grained taxonomy. It identifies three trends: limited real-world scalability, decreasing supervision requirements, and benefits from combining ideas across method clusters.

Key findings

• Across four datasets, reduced supervised methods approach supervised performance: on CIFAR-10, MixMatch and FixMatch exceed 95% accuracy and are roughly 3% below the fully supervised baseline; on CIFAR-100, FixMatch reaches about 77% versus a supervised baseline of about 80%.
• On STL-10, EnAET achieves more than 95%, and FixMatch reports nearly 95% with only 1,000 labels; reduced supervised methods can surpass the supervised baseline by about 20% due to the additional unlabeled data.
• On ILSVRC-2012, SimCLRv2 achieves the best reported Top-1 accuracy of 80.9% and Top-5 accuracy around 96%; only SimCLR, BYOL, and SimCLRv2 are less than 10% worse than the supervised baseline of around 86%.
• The survey concludes that reduced supervised learning reaches comparable results using roughly 10% of labels, and that combining ideas from different clusters can improve performance.

Methods

• Survey design: compares 34 methods for image classification, grouping them by training strategy—one-stage-semi-supervised, one-stage-unsupervised, and multi-stage-semi-supervised—and by common ideas such as cross-entropy, pseudo-labels, pretext tasks, contrastive loss, mutual information, overclustering, and mixup.
• Data and evaluation: analyzes reported results on CIFAR-10, CIFAR-100, STL-10, and ILSVRC-2012, using classification accuracy or cluster accuracy for unsupervised methods.
• Analysis: relies on results reported in the original papers or cross-references rather than recreating every method in a unified setup, and focuses on general trends and specialties instead of small performance differences.

Limitations

• Stated: the comparison uses reported values rather than a unified reimplementation, and differences in architectures, implementations, and adherence to semi-supervised evaluation guidelines may affect comparability.
• Stated: the common-idea framework is fuzzy and incomplete by definition, and the survey excludes related topics such as metric learning, meta learning, GANs, graph networks, pose estimation, segmentation, videos, sketches, and few-/zero-shot learning.
• Evident: the survey does not claim a complete method list and does not present new experiments; its conclusions depend on the availability and consistency of results reported by the reviewed papers.