instance discrimination

**Instance Discrimination** is the **foundational contrastive learning paradigm where each image in the dataset is treated as its own unique class** — and the model is trained to distinguish each instance from all others, learning representations that capture fine-grained visual differences. **What Is Instance Discrimination?** - **Definition**: Treat the N images in the dataset as N classes. - **Positive**: Augmented versions of the same image. - **Negative**: All other images. - **Loss**: NCE/InfoNCE applied to the N-class discrimination task. - **Paper**: Wu et al., "Unsupervised Feature Learning via Non-Parametric Instance Discrimination" (2018). **Why It Matters** - **Foundation**: SimCLR, MoCo, BYOL, and DINO are all built on the instance discrimination framework. - **No Labels Needed**: The "class" of each image is its identity — no human annotation required. - **Semantic Emergence**: Despite training with instance-level labels, learned features capture semantic similarity (a surprising and powerful property). **Instance Discrimination** is **the philosophical foundation of contrastive SSL** — the insight that treating every image as unique can paradoxically teach a model to understand what makes images similar.