Source
Caron et al. (2020), LIT-598 — ARXIV-2006.09882.
The claim
A joint-embedding method compares two augmented views. More views would help and two is what the budget allows — but only because everyone assumed the views must be the same size.
Use a mix of resolutions. Two standard-resolution crops plus several smaller ones. The number of comparisons rises; the compute and memory do not, because the extra views are small. Nothing else about the method changes.
Why this is unusually well-evidenced for its size
Multi-crop was proposed in LIT-598, Caron et al.’s SwAV paper, as a change to the augmentation separable from SwAV’s swapped-prediction objective: two standard-resolution crops plus several smaller ones, so the number of views rises without a matching rise in compute or memory.
The paper that introduced it did not keep it. Multi-crop was applied to SimCLR, DeepCluster and DeepCluster-v2 — three methods with three different objectives, one of them contrastive with negatives and two clustering-based — and “consistently improves the performance for all the considered methods by a significant margin of 2–4% top-1”.
That is the transplant test the record usually has to wait for, performed in the introducing paper. Compare SOTA-367’s variance term, the only other component in this cluster with evidence of working outside its own method.
What it says about the mechanism, which is more than it looks
If view count is what matters and view resolution is mostly what costs, then the thing a joint-embedding loss is learning from is the number of independent comparisons, not the fidelity of each one. That is consistent with SOTA-361 — the augmentation set specifies the task — and it sharpens it: the task is specified by the distribution of views, and you can sample it more densely for free.
Conditions
- The small crops are not free of assumptions. A low-resolution crop of an image is still recognisable as the same object; a low-resolution crop of a text span, a spectrogram or a point cloud may not be. The resolution/semantics trade is domain-specific and the paper only tests it on natural images.
- Measured on ImageNet, on four methods. Broad for a component result, narrow in domain.
- It compounds with the augmentation policy rather than replacing it (SOTA-361). The crops still need the colour distortion.
- The asymmetry it introduces is real: with views of different sizes the two branches no longer see the same distribution, which interacts with whatever anti-collapse mechanism is in use. SwAV’s own is the equipartition constraint, not an architectural asymmetry, so the interaction is untested for the predictor-and-stop-gradient family (SOTA-365).