Research map / Label-efficient and robust learning / Semi-supervised learning
Info-Coevolution: An Efficient Framework for Data Model Coevolution
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 Info-Coevolution, an online selective-annotation framework that estimates a sample's information gain via model predictions fused with kNN data-view predictions (Bayesian fusion), cutting ImageNet-1K annotation to 68% without loss and to 50% with semi-supervised learning.
Problem
Deciding whether new data samples or batches need annotation or training given the current model and dataset, in growing real-world data streams, while avoiding the pipeline complexity and bias of active learning.
Approach
Estimates a sample's annotation gain by fusing the model's confidence with a kNN prediction from the already-annotated data in feature space (Bayesian prediction fusion), with dynamic rechecking of neighbors via online ANN (HNSW), so the model is updated only infrequently.
Key results
- On ImageNet-1K, Info-Coevolution reaches lossless accuracy (85.6%) with only 68% of annotations, versus Semi-ViT's 85.5% at 80% labels.
- At 10% ImageNet labels, selective annotation reaches 80.2% (80.5% with two model updates) vs 79.2% for random selection.
- With Semi-ViT and 50% of ImageNet labels selected by Info-Coevolution, accuracy is 85.5%, described as almost lossless.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use prioritized Info-Coevolution selection instead of random selection | ViT-Large supervised training · ImageNet-1K, 10% labels | accuracy: +1.3% | Helped |
| Use Info-Coevolution selection instead of random selection | ResNet-50x4 with Fixmatch · CIFAR10, 250 selected labels | accuracy: 94.95 -> 95.39 | Helped |
| Use Info-Coevolution selection instead of random selection | ResNet-50x4 with Fixmatch · CIFAR10, 1000 selected labels | accuracy: 95.59 -> 95.85 | Helped |
| Add unlabeled data to semi-supervised training | Semi-ViT · ImageNet-1K, 1.28M labeled data | accuracy: 85.6 -> 86.0 | Helped |
| Select 1M unlabeled samples with Info-Coevolution rather than use 2M unlabeled samples | Semi-ViT · ImageNet-1K, 1.28M labeled data | accuracy: 86.0 -> 86.0 | No clear effect |
| Use model prediction only for sample-gain estimation (ablation) | Info-Coevolution selection · ImageNet-1K, 10% labels | accuracy: 79.2 -> 78.5 | Hurt |
| Use data-view prediction for sample-gain estimation (ablation) | Info-Coevolution selection · ImageNet-1K, 10% labels | accuracy: +0.6% | Helped |
| Add dynamic rechecking to model-only sample selection (ablation) | Info-Coevolution selection · ImageNet-1K, 10% labels | accuracy: +0.5% | Helped |
| Fuse model and KNN predictions without dynamic rechecking (ablation) | Info-Coevolution selection · ImageNet-1K, 10% labels | accuracy: -2.1% | Hurt |
| Add dynamic rechecking to fused model and KNN predictions (ablation) | Info-Coevolution selection · ImageNet-1K, 10% labels | accuracy: +1.0% | Helped |
| Add two model updates during selection (ablation) | Info-Coevolution · ImageNet-1K, 10% labels | accuracy: 80.2% -> 80.5% | Helped |
Surprising findings
- Using only the model's prediction for selection fell below the random baseline (78.5% vs 79.2%), and fused prediction alone caused a class-balance problem (77.1%).
Limitations
- Assumes the same training distribution as the target; weakly supervised web data may differ in distribution.
- Unlabeled and weakly supervised extensions are only preliminarily studied.
- Evaluated only on classification tasks; other tasks are left to future work.
Experimental setup
Datasets: ImageNet-1K CIFAR-10 CIFAR-100 StanfordCars Food-101 SVHN CC3M CC12M SBU COCO Visual Genome LAION-400M
Benchmarks: ImageNet-1K CIFAR-10 CIFAR-100 StanfordCars Food-101 SVHN
Metrics: top-1 accuracy annotation ratio
Baselines: Random selection Semi-ViT Dataset Quantization FixMatch
Builds on: Semi-ViT Dataset Quantization HNSW FixMatch MAE
Closest papers by meaning
- InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning (ICLR 2024 · similarity 0.84)
- DataRater: Meta-Learned Dataset Curation (NeurIPS 2025 · similarity 0.83)
- Scaling Laws for Data Filtering-- Data Curation cannot be Compute Agnostic (CVPR 2024 · similarity 0.83)
- When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration (ICML 2025 · similarity 0.83)
- Data-Efficient Learning via Clustering-Based Sensitivity Sampling: Foundation Models and Beyond (ICML 2024 · similarity 0.83)
- CoLAL: Co-learning Active Learning for Text Classification (AAAI 2024 · similarity 0.83)
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.