RotNet vs. Exemplar: Stability Issues in Early Self-Supervised Learning

A_K_Nain · x · 2026-08-25

The post revisits self-supervised learning (SSL) history. RotNet (2018) predicts image rotations to learn features, appearing simple and effective. However, a comparative study found that, like other 'creative' SSL proxy tasks, RotNet's performance varies significantly with minor architecture changes. In contrast, the Exemplar method remained stable and is arguably the true predecessor to modern contrastive methods like SimCLR and MoCo.

Original post →

More from Research

Research channel →