meta-learning

**Meta-Learning Few-Shot Learning** is **training systems to quickly learn new tasks from few examples, mimicking human ability to generalize from limited data through learned inductive biases** — enables rapid adaptation. Meta-learning learns to learn. **Few-Shot Learning Problem** train on diverse tasks with few examples per task. Test on new task with few examples. Goal: learn from little data. **Task Distribution** different tasks sampled from task distribution. Meta-training: learn across tasks. Meta-testing: adapt to new task. **Model-Agnostic Meta-Learning (MAML)** gradient-based meta-learning: learn initial parameters enabling fast adaptation. Inner loop: gradient step(s) on new task. Outer loop: optimize for few-shot performance. **Meta-Gradient** gradient of gradient. Compute gradient for new task, then gradient of that loss at new points. Second-order derivatives. **Prototypical Networks** metric learning: embed examples in space, novel class centroid (prototype) is mean embedding of few examples. Classify by nearest prototype. **Matching Networks** attention-based: compute attention weights over support set examples, predict class via attention-weighted sum. Similar to prototypical networks. **Relation Networks** learn similarity metric instead of assuming Euclidean distance. Neural network predicts relation score between query and support examples. **Optimization-Based Meta-Learning** MAML, learned optimizers. Learn parameters enabling fast gradient descent. **Metric-Based Meta-Learning** prototypical networks, matching networks, relation networks. Learn embeddings/similarity. **Siamese Networks** pairs of inputs: same class (positive) vs. different class (negative). Contrastive loss. Learn discriminative embeddings. **Memory-Augmented Networks** external memory for rapid adaptation. Attention over memory stores learned knowledge. Neural Turing Machines. **Embedding Learning** learn good representation space where few examples suffice for classification. Representation transfer. **Data Augmentation for Few-Shot** augment few examples generating synthetic examples. Mixup, style transfer. **Transfer Learning vs. Meta-Learning** transfer: pretrain on source, finetune on target. Meta-learning: learn to finetune. Different philosophy. **N-Way K-Shot** N classes, K examples per class (few-shot). Standard evaluation: 5-way 5-shot. **Benchmark Datasets** omniglot (handwritten characters), miniImageNet, CUB (birds), Caltech-256. **Cross-Domain Few-Shot** train on one domain, test on another. Harder: significant distribution shift. **Zero-Shot Learning** no examples of new class. Use semantic attributes or word embeddings. Extreme generalization. **Task Augmentation** generate synthetic tasks for meta-training. Improve meta-learning. **Episodic Training** organize meta-training as episodes (tasks). Sample support/query sets each episode. Better matches meta-test. **Uncertainty in Few-Shot** Bayesian few-shot learning: posterior over parameters given few examples. **Long-Tail Distribution** many classes with few examples. Meta-learning naturally applicable. **Domain Generalization** meta-learning improves out-of-distribution generalization. Learning across diverse tasks. **Multi-Task Meta-Learning** meta-learn across multiple related meta-tasks. **Applications** robotics (quickly adapt to new environment), natural language (few-shot text classification), computer vision (few-shot object detection). **Meta-Learning Frameworks** learn2learn, higher libraries simplify meta-learning. **Theoretical Analysis** meta-learning convergence, sample complexity. **Few-Shot Meta-Learning enables rapid adaptation to new tasks** from minimal data, approaching human generalization.