Research map / Label-efficient and robust learning / Semi-supervised learning
Reinforcement Learning Guided Semi-Supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formulates semi-supervised learning as a one-armed bandit and trains a classifier with a KL-weighted mixup reward RL loss, a teacher-student EMA framework, and supervised and consistency losses.
Problem
Semi-supervised image classification with scarce labels, where existing SSL methods rely on heuristic pseudo-labels and predefined loss/regularization designs.
Approach
Treats the classifier as a probabilistic policy whose soft pseudo-label outputs on unlabeled data are actions, and rewards them by negative MSE on mixup (labeled/unlabeled) interpolations; the policy loss is a KL-from-uniform weighted negative reward, with actions produced by an EMA teacher.
Key results
- On CIFAR-10 with 1000 labels (CNN-13), RLGSSL reaches 9.15% test error vs 12.44% for ICT.
- On CIFAR-100 with 10000 labels (CNN-13), RLGSSL reaches 29.12% vs 32.24% for ICT.
- On SVHN with 1000 labels (WRN-28), RLGSSL reaches 1.92% vs 1.99% for Meta Pseudo-Labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RLGSSL instead of ICT | CNN-13 · CIFAR-10/1000 labels | test error: 12.44 -> 9.15 | Helped |
| use RLGSSL instead of ICT | CNN-13 · CIFAR-10/2000 and 4000 labels | test error: 8.69 -> 6.90; 7.18 -> 6.11 | Helped |
| use RLGSSL instead of ICT | CNN-13 · CIFAR-100/4000 and 10000 labels | test error: 40.07 -> 36.92; 32.24 -> 29.12 | Helped |
| use RLGSSL instead of SNTG | CNN-13 · SVHN/500 and 1000 labels | test error: 3.99 -> 3.12; 3.86 -> 3.05 | Helped |
| use RLGSSL instead of MixMatch | WRN-28 · CIFAR-10/1000, 2000, and 4000 labels | test error: 7.75 -> 4.92; 7.03 -> 4.24; 6.24 -> 3.52 | Helped |
| remove RL loss (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 44.92 | Hurt |
| remove supervised loss (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 39.52 | Hurt |
| remove consistency loss (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 38.78 | Hurt |
| remove EMA teacher update (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 43.12 | Hurt |
| remove mixup (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 40.12 | Hurt |
| set reward to constant 1 (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 39.52 | Hurt |
| set μ=0 so reward uses labeled data only (ablation) | RLGSSL · CIFAR-100/4000 labels | test error: 36.92 -> 39.54 | Hurt |
Limitations
- Reward and RL loss are heuristic designs; the paper gives no theoretical justification that the bandit formulation improves SSL.
- Gains over ICT on CIFAR-10/100 are reported with few baselines reproduced under identical settings; hyperparameter sensitivity is only shown on one CIFAR-100 setting.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 SVHN
Benchmarks: CIFAR-10 CIFAR-100 SVHN
Metrics: test error (%) standard deviation over 5 trials
Baselines: Supervised + MixUp Pi-model Temporal Ensembling Mean Teacher VAT SNTG Learning to Reweight MT + Fast SWA ICT MixMatch Meta Pseudo-Labels
Builds on: Mean Teacher mixup Virtual Adversarial Training Interpolation Consistency Training single-step MDP
Closest papers by meaning
- TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning (ICLR 2026 · similarity 0.88)
- Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning (ICLR 2026 · similarity 0.88)
- Can RLHF be More Efficient with Imperfect Reward Models? A Policy Coverage Perspective (ICML 2025 · similarity 0.87)
- Critic-Guided Decision Transformer for Offline Reinforcement Learning (AAAI 2024 · similarity 0.87)
- Gradient Boosting Reinforcement Learning (ICML 2025 · similarity 0.87)
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.