Position: Unlabeled IS NOT Equal to No Human Supervision in Visual Learning
Dong Lao
ICML 2026
cs.CV, cs.LG
2026-09-03
LSU's ICML paper finds CVPR unsupervised titles fell from ~4% in 2021 to ~1% in 2025; unlabeled training still embeds human priors that should be disclosed.
Self-supervised pre-training turned visual features into infrastructure. Downstream papers that never touch human labels still get filed as unsupervised. That is true about labels and incomplete about supervision. Curation, objectives, and which backbone you inherit already bake in human priors.
Dong Lao at LSU argues in this ICML 2026 position paper that missing labels is not missing human supervision. Earlier unsupervised work was mostly in-domain: digits on digits, faces on faces. Those distributional bets were local and visible. Foundation-scale pre-training ships the same bets across domains, and they rarely get named.
This is a community audit plus a disclosure checklist, not a new trainer.
Title counts cover CVPR, ICCV, and ECCV main papers from 2013 to 2025, tagging unsupervised and self-supervised while excluding unsupervised domain adaptation, where shift is the object of study. A full-text pass over all 17,435 CVPR papers from 2015 to 2025 uses Qwen-2.5-32B-Instruct to record only author-stated pre-training dependence. For the 287 CVPR papers from 2020 to 2025 with unsupervised in the title, humans checked whether the method relies on a pre-trained model outside the paper's core claim. Against that annotation the automatic classifier hits 91.5% precision, 81.8% recall, and 86.9% accuracy.
Supervision enters through two cracks. Unfiltered Google Image results for tiger and cup already show canonical poses, centered crops, and cup handles facing right; that is photographer and search habit, not an emergent algorithm. Contrastive learning assumes two augmentations of one image are the same entity. That holds on object-centric ImageNet shots and breaks in multi-object scenes or microscopy. DINO-family models improve with scale on object-centric data and drop on ImageNet-1k linear eval once generic, non-object-centric images are mixed in. Horizontal flip is a fair invariance for generic objects and a semantic inversion for road signs, so the invariance itself is a human call.
The checklist does not penalize priors. It asks authors to state whether pre-training is used, which layers are frozen, data scope and filters, enforced invariances, and evaluation leakage. Companion practices include splitting pre-training-dependent pipelines from in-domain unlabeled learners, testing across contrastive, reconstruction, and multimodal backbones, treating assumption relaxation as a contribution, and comparing at similar compute.
The numbers describe how the community talks, not a new SOTA.
| signal | peak | later |
| CVPR titles with unsupervised | nearly 4% in 2021 | close to 1% by 2025 |
| CVPR full text on un/self-supervised | 22.17% in 2021 (10.13% in 2015) | 15.52% in 2024, 12.85% in 2025 |
Among CVPR papers titled unsupervised, dependence on external pre-training has risen since 2020 and concentrates in DINO, CLIP, MoCo, diffusion models, and SwAV. A few backbone families account for most named sources, so downstream "emergence" may be shared upstream curation.
Three hypotheses sit side by side, without causal identification: weaker-assumption methods look worse under one unsupervised umbrella, and scaling data or compute pays off faster than a new objective; once pre-training is infrastructure, unsupervised stops being the claimed contribution in the title; end users care about plugging a model in more than about the training recipe.
Practitioners freeze DINO or CLIP every week and call the pipeline unsupervised segmentation, discovery, or correspondence. If the real work was done by upstream curation and an objectness prior, writing it up as downstream emergence makes reproduction and transfer fragile. A short checklist is cheap and puts two unsupervised papers on comparable terms.
For groups without industrial GPUs the stakes are higher. If unsupervised quietly means you can afford a giant backbone, in-domain, weak-assumption, motion, or geometry work gets harder to publish. Treating explicit assumptions and relaxed assumptions as publishable contributions keeps that work in the room.
This is a position, not a method. It changes how papers are written and compared, not which pre-training recipe you should run tomorrow.
Title statistics are coarse, and the paper says so. Dropping domain adaptation is defensible and still misses work that never puts unsupervised in the title. The full-text scan uses Qwen-2.5-32B and only records named pre-training, so both misses and false hits remain. The 287-paper manual set is a title-selected slice, not all of vision.
The three hypotheses are not causally tested. The DINO scale plot shows an objective-data mismatch but does not publish a citable absolute accuracy. Whether the checklist becomes a review ritual or empty boilerplate is untested. By refusing to name specific unsupervised papers, the argument stays cleaner and harder to apply to any one method.