one-shot learning

**One-shot learning** is the extreme case of few-shot learning where a model must learn to recognize or classify new categories from **just a single example per class**. This mirrors human cognitive abilities — people can often identify a new object after seeing it only once by leveraging extensive prior knowledge. **Why One-Shot is Especially Challenging** - **Single Point Representation**: With only one example, any noise, unusual angle, or atypical instance creates a skewed class representation. - **No Variance Estimation**: Cannot estimate intra-class variability from a single example — the model doesn't know what range of appearances to expect. - **Overfitting Risk**: Standard fine-tuning on one example leads to immediate overfitting. **Technical Approaches** - **Siamese Networks**: Learn a **similarity function** that compares input pairs and determines whether they belong to the same class. Uses **contrastive loss** or **triplet loss** to train discriminative embeddings. - Input: Two images → Output: Same class or different class (with confidence). - At test time: Compare the query against the single reference example. - **Matching Networks**: Use an **attention mechanism** over the support set to classify queries based on learned similarity kernels. The full context of the support set influences each classification decision. - **Memory-Augmented Neural Networks (MANN)**: Store examples in a **differentiable external memory** and retrieve relevant stored examples for new queries. Enables rapid binding of new information without modifying network weights. - **Prototypical Networks**: With K=1, the prototype is simply the single example's embedding. Classification relies entirely on the quality of the learned embedding space. **Key Benchmarks** - **Omniglot**: 1,623 handwritten characters from 50 different alphabets, each drawn by 20 people. A "transpose" of MNIST — many classes, few examples. Standard 5-way 1-shot accuracy: ~98%. - **miniImageNet**: 5-way 1-shot accuracy for state-of-the-art methods: ~65–75% (much harder than Omniglot). - **CUB-200 Birds**: Fine-grained one-shot species identification. **Modern Approaches** - **Large Pre-Trained Models**: Vision-language models like **CLIP** and **DINOv2** provide rich feature representations that enable effective one-shot transfer. CLIP can even perform **zero-shot** classification through natural language class descriptions. - **Data Augmentation**: Apply aggressive augmentations to the single example — rotations, crops, color jitter, CutMix — to artificially increase the training signal. - **Hallucination Networks**: Generate synthetic additional examples from the single reference using learned transformations. **Applications** - **Face Recognition**: Identify individuals from a single enrollment photo (security, access control). - **Signature Verification**: Authenticate signatures from a single genuine reference. - **Drug Discovery**: Screen compounds based on single known active molecule structures. - **Robotics**: Recognize new objects or tools from a single demonstration. One-shot learning represents the **frontier of data-efficient AI** — it pushes the limits of how much a model can learn from minimal data, a capability essential for deploying AI in data-scarce environments.