exemplar learning

**Exemplar learning** is the **early self-supervised approach that groups multiple augmentations of the same image into one pseudo-class to learn invariant features** - it predated large-scale contrastive pipelines and demonstrated that transformation consistency can supervise representation learning. **What Is Exemplar Learning?** - **Definition**: Generate transformed variants of each image and train network to treat those variants as related exemplars. - **Pseudo-Label Strategy**: Each source image forms a pseudo category under augmentation. - **Objective Choices**: Triplet loss, pairwise metric losses, or proxy classification variants. - **Historical Context**: Important stepping stone toward modern instance contrastive methods. **Why Exemplar Learning Matters** - **Invariance Learning**: Encourages robustness to rotation, crop, color, and geometric transformations. - **Label-Free Supervision**: Uses synthetic relationships without manual annotation. - **Method Simplicity**: Clear augmentation-driven supervisory signal. - **Legacy Influence**: Inspired later methods that formalized positive-pair construction. - **Educational Value**: Useful baseline for understanding SSL objective evolution. **How Exemplar Learning Works** **Step 1**: - Apply multiple stochastic augmentations to each image to create exemplar set. - Encode exemplars into embedding space with shared backbone. **Step 2**: - Optimize metric objective so exemplars from same source are close and others remain separated. - Repeat across dataset to build transformation-invariant representation geometry. **Practical Guidance** - **Augmentation Diversity**: Too weak gives poor invariance, too strong can remove semantics. - **Triplet Sampling**: Hard negative mining often improves convergence quality. - **Scale Limits**: Large pseudo-class counts can stress memory and classifier design. Exemplar learning is **an early but influential SSL strategy that proved augmentation consistency can replace manual labels for representation training** - it remains a useful conceptual baseline for modern self-supervised pipelines.