Self-supervised learning often relies on augmentations to create multiple views of the same input, then learns representations that capture what is shared across those views. But as a new arXiv preprint notes, this shared content is typically recovered as a whole, without separating the underlying factors into individual coordinates. The authors ask whether the choice of augmentation can be designed to change that.
The title and abstract point to 'variance fingerprints' as a mechanism for component-wise identification of content factors. However, the available abstract is truncated, so the full method and experimental results are not described in the source. The preprint is listed as arXiv:2610.14260.