Research map / Label-efficient and robust learning / Noisy labels
CoLAL: Co-learning Active Learning for Text Classification
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.
Proposes CoLAL, an active learning method that uses a peer model trained on the target model's noisy predictions on unlabeled data, selecting samples where the two models disagree most, and reports gains over text AL baselines.
Problem
Active learning under limited labeled data: selecting unlabeled samples that are diverse, representative, and likely to be misclassified by the target model near decision boundaries.
Approach
A peer model is trained with noisy labels (target model predictions) on unlabeled data for only 2 epochs to exploit early memorization; a probabilistic graphical model turns the multi-class agreement into a binary same-prediction classifier, and samples where the target and peer models disagree are selected by a disagreement score (1 minus the inner product of their categorical distributions).
Key results
- On text benchmarks, CoLAL outperforms the best baseline by average gaps of about 0.6% to 2.0% on AGNews and 1.0% to 2.2% on SST-2 across sample sizes, with statistically significant improvements per the authors' t-tests.
- On PubMed, CoLAL beats the best baseline BADGE by 1.955% on average.
- On image datasets, CoLAL is competitive with AlphaMix, matching it by the 5th iteration on MNIST and slightly outperforming baselines in iterations 6-10 on OpenML-155.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoLAL instead of the strongest baseline | Text classification · AGNews | accuracy: average gaps ranging from 0.642% to 2.050%; peak differences between 1.103% and 3.26% at 800 samples | Helped |
| Use CoLAL instead of the strongest baseline | Text classification · DBPedia | accuracy: average accuracy gaps spanning from 0.543% to 1.733%; highest outperformance between 1.646% and 2.391% at 200 and 400 sa | Helped |
| Use CoLAL instead of the strongest baseline | Text classification · PubMed | accuracy: average accuracy improvements ranging from 1.026% to 2.046%; against the best baseline BADGE at 1.955% | Helped |
| Use CoLAL instead of the strongest baseline | Text classification · SST-2 | accuracy: average accuracy gaps over other methods between 1.026% and 2.219%; maximum improvements observed at 3.628% against BADG | Helped |
| Use CoLAL instead of image-classification baselines | Image classification · CIFAR100 | accuracy | Mixed |
| Use CoLAL instead of AlphaMix | Image classification · MNIST | accuracy: 5th iteration | No clear effect |
| Use CoLAL instead of image-classification baselines | Image classification · OpenML-155 | accuracy: iterations 6 to 10 | Helped |
Surprising findings
- The peer model is intentionally trained for only 2 epochs on noisy labels, relying on early-stage memorization of clean patterns rather than long training.
Limitations
- Peer model epoch count (1 vs 2) was selected empirically for BERT-based embeddings, limiting generality.
- Image-classification results are only competitive with, not clearly superior to, AlphaMix.
Experimental setup
Datasets: AGNews DBPedia PubMed SST-2 CIFAR100 MNIST OpenML-155
Benchmarks: AGNews DBPedia PubMed SST-2 CIFAR100 MNIST OpenML-155
Metrics: accuracy
Baselines: Random Max-Entropy BADGE CAL AcTune BATL AlphaMix
Builds on: co-teaching DAL CAL AlphaMix
Closest papers by meaning
- Fast Co-Training under Weak Dependence via Stream-Based Active Learning (ICML 2024 · similarity 0.86)
- A Cross-Domain Benchmark for Active Learning (NeurIPS 2024 · similarity 0.86)
- Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning (CVPR 2025 · similarity 0.86)
- Composite Active Learning: Towards Multi-Domain Active Learning with Theoretical Guarantees (AAAI 2024 · similarity 0.85)
- Instance-wise Supervision-level Optimization in Active Learning (CVPR 2025 · similarity 0.84)
- Contextual Active Model Selection (NeurIPS 2024 · similarity 0.84)
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.