Research map / Vision-language and multimodal / Multimodal LLMs
Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding
Talfan Evans, Shreya Pathak, Hamza Merzic, Jonathan Richard Schwarz, Ryutaro Tanno, Olivier Henaff
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| prioritize high-loss data under the learner (hard learner) (ablation) | ViT-B on JFT-300M · JFT-300M | accuracy / speedup | No clear effect |
| add a pretrained reference model (easy-reference or learnability scoring) (ablation) | ViT-B on JFT-300M · JFT-300M | learner speedup: speed-ups of ~33% | Improved |
| use learnability instead of easy-reference scoring with small scoring models (ViT-S, ViT-Ti) (ablation) | ViT-L on JFT-300M · JFT-300M | learner speedup: 31% (ViT-B) -> 26% (ViT-Ti) | Improved |
| scale down scoring models to ViT-Mu family (about 1000x smaller) (ablation) | ClassAct · JFT-300M, ViT-L | learner speedup: 16% for the smallest actors | Mixed |
| RHO with small reference model (ablation) | ViT-B learner · JFT half, 3 epochs | learner speedup: 0% learner speedup, - 79% compute speedup | Hurt |
| add small online model to scoring (ClassAct) instead of using the learner for scoring (ablation) | RHO · JFT half, 3 epochs | compute speedup: - 79% -> 3% | Improved |
| train reference model online in parallel on larger super-batch (ablation) | ClassAct · JFT half, 3 epochs | learner speedup: 18% -> 17% | No clear effect |
| reference model trained on held-out set instead of in-domain data (ablation) | ClassAct · JFT half, 3 epochs | learner speedup: 18% -> 18% | No clear effect |
| use LTIP (cleaner) reference model instead of ALIGN reference (ablation) | ActiveCLIP · ImageNet 0-shot, ALIGN learner | top-1 accuracy: 47.8 -> 53.2 | Improved |
| use ViT-Ti policy instead of ViT-B policy for ActiveCLIP (ablation) | ActiveCLIP · ImageNet 0-shot | top-1 accuracy / FLOP saving: 70.3 -> 69.9; FLOP saving -163.4% -> 26.2% | Mixed |
| reference model trained on 350M curated Webli subset (ablation) | ActiveSigLIP · Webli | transfer performance | Improved |
| apply ActiveSigLIP on top of SigLIP | SigLIP · COCO retrieval and ImageNet 0-shot | COCO im2txt R@1: 60.7 -> 63.5 (im2txt), 42.7 -> 45.3 (txt2im), ImageNet 72.1 -> 72.0 | Mixed |
Closest papers by meaning
- Learn How to See: Collaborative Embodied Learning for Object Detection and Camera Adjusting (AAAI 2024 · similarity 0.81)
- From EduVisBench to EduVisAgent: A Benchmark and Multi-Agent Framework for Reasoning-Driven Pedagogical Visualization (ICLR 2026 · similarity 0.81)
- Attention Transfer Is Not Universally Effective for Vision Transformers (NeurIPS 2026 · similarity 0.80)
- Masking Teacher and Reinforcing Student for Distilling Vision-Language Models (CVPR 2026 · similarity 0.80)
- Multi-Teacher Knowledge Distillation with Reinforcement Learning for Visual Recognition (AAAI 2025 · similarity 0.80)
- DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026 · similarity 0.79)
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.
