Research map / Label-efficient and robust learning / Recognition and long-tail
Dense features and foundation priors
Learns dense or tabular features using synthetic priors, hierarchical contrast and feature re-embedding.
482 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “representations · transfer · unsupervised”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
148 papers in 2023–24 and 196 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 11.8% to 14.9%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Column Sampling Based Discrete Supervised Hashing (AAAI 2016 · 314 citations)
- 2017 · Context-Aware Correlation Filter Tracking (CVPR 2017 · 666 citations)
- 2018 · TieNet: Text-Image Embedding Network for Common Thorax Disease Classification and Reporting in Chest X-Rays (CVPR 2018 · 537 citations)
- 2019 · Transfusion: Understanding Transfer Learning for Medical Imaging (NeurIPS 2019 · 637 citations)
- 2020 · Neural Cognitive Diagnosis for Intelligent Education Systems (AAAI 2020 · 314 citations)
- 2021 · Dual-GAN: Joint BVP and Noise Modeling for Remote Physiological Measurement (CVPR 2021 · 196 citations)
- 2022 · InfoGCN: Representation Learning for Human Skeleton-Based Action Recognition (CVPR 2022 · 442 citations)
- 2023 · V2X-Seq: A Large-Scale Sequential Dataset for Vehicle-Infrastructure Cooperative Perception and Forecasting (CVPR 2023 · 176 citations)
- 2024 · Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology (CVPR 2024 · 104 citations)
- 2025 · SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · 56 citations)
- 2026 · DriveSuprim: Towards Precise Trajectory Selection for End-to-End Planning (AAAI 2026 · 4 citations)
Earliest papers here
- Column Sampling Based Discrete Supervised Hashing (AAAI 2016 · 314 citations)
- Learning Deep Representation from Big and Heterogeneous Data for Traffic Accident Inference (AAAI 2016 · 229 citations)
- Instilling Social to Physical: Co-Regularized Heterogeneous Transfer Learning (AAAI 2016 · 48 citations)
What has been tried
344 papers here report 3,465 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 7%, hurt 19%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add R2T feature re-embedding to AB-MIL with ResNet-50 features | CAMELYON-16 · AUC | 94.52 -> 97.32 | Helped | Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology (CVPR 2024 · 104 citations) |
| Add expert feature distillation to TCP-traj | Bench2Drive base · Driving Score | 49.30 -> 59.90 | Helped | Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving (NeurIPS 2024 · 81 citations) |
| adding ego status to the BEV-Planner (BEV-Planner++) | nuScenes open-loop planning · avg L2 (m) | 0.55 -> 0.35 | Helped | Is Ego Status All You Need for Open-Loop End-to-End Autonomous Driving? (CVPR 2024 · 78 citations) |
| Add Pad-and-Resize | CCPG · rank-1 accuracy (CL) | 70.5 -> 76.0 | Helped | BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · 58 citations) |
| Use SatCLIP L=40 instead of L=10 | Biome · accuracy | 92.23 ± 0.26 -> 94.41 ± 0.14 | Helped | SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · 56 citations) |
| replace interpolation with adaptive max pooling | NSD subject 1 · CLIP | 94.4% -> 95.0% | Helped | MindBridge: A Cross-Subject Brain Decoding Framework (CVPR 2024 · 46 citations) |
| Add velocity and acceleration to goal-only ego status | navtest · PDMS | 81.8 -> 84.0 | Helped | NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking (NeurIPS 2024 · 43 citations) |
| add mixed data augmentation | SEED · accuracy | 85.42 -> 88.27 | Helped | DMMR: Cross-Subject Domain Generalization for EEG-Based Emotion Recognition via Denoising Mixed Mutual Reconstruction (AAAI 2024 · 42 citations) |
| Set λ=1 instead of λ=0 in the hierarchical probability loss | MA-52 · action F1_macro | 52.69 -> 53.83 | Helped | Prototypical Calibrating Ambiguous Samples for Micro-Action Recognition (AAAI 2025 · 38 citations) |
| Add the target encoding module (TEM) | SCC · PCC(M) | 0.352 -> 0.374 | Helped | Accurate Spatial Gene Expression Prediction by Integrating Multi-Resolution Features (CVPR 2024 · 37 citations) |
| Use four training observation lengths (10, 16, 20, 30) rather than three (10, 20, 30) | Argoverse 1 validation, 10/20/30 timesteps · ADE6 | 0.81 -> 0.79; 0.72 -> 0.72; 0.65 -> 0.68 | Mixed | Adapting to Length Shift: FlexiLength Network for Trajectory Prediction (CVPR 2024 · 27 citations) |
| add Instance-level Expert Knowledge (IEK) to the BASE model | IU X-Ray · BLEU-1 | 0.463 -> 0.501 | Helped | Instance-level Expert Knowledge and Aggregate Discriminative Attention for Radiology Report Generation (CVPR 2024 · 25 citations) |
What reviewers will expect
Datasets and baselines named in the 110 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- nuScenes · 7 of 110 papers
- Bench2Drive · 4 of 110 papers
- MNIST · 4 of 110 papers
- ImageNet · 3 of 110 papers
- Fashion MNIST · 3 of 110 papers
- CIFAR100 · 3 of 110 papers
- CSI300 · 2 of 110 papers
- CSI500 · 2 of 110 papers
Baselines
- UniAD · 6 of 110 papers
- XGBoost · 5 of 110 papers
- VAD · 5 of 110 papers
- Random Forest · 4 of 110 papers
- LightGBM · 4 of 110 papers
- SparseDrive · 4 of 110 papers
- MLP · 3 of 110 papers
- K-means · 3 of 110 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 344 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 18 papers
- Tsinghua University · 13 papers
- Shanghai Jiao Tong University · 9 papers
- Massachusetts Institute of Technology · 8 papers
- University of Science and Technology of China · 8 papers
- University of Chinese Academy of Sciences · 7 papers
- Stanford University · 7 papers
- Harbin Institute of Technology · 7 papers
- University of Oxford · 6 papers
- Carnegie Mellon University · 6 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Contrastive representation theory (495 papers)
- Correspondence and causal representation learning (447 papers)
- Data augmentation and generalization bias (409 papers)
- Multilingual language representation pretraining (507 papers)
- Recommendation representation learning (207 papers)
Elsewhere in the topic
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
- Learning from unlabeled and positive-unlabeled data (240 papers)
- Confidence calibration objectives (129 papers)
- Margin-based face recognition losses (124 papers)
- Supervised person re-identification (108 papers)
- Self-explaining and visual explanation networks (108 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).