Matching Networks compare query examples to support set using attention mechanism for few-shot classification. Approach: Learn embeddings and attention-based comparison. Query attends to all support examples, weighted combination determines class. Architecture: Embedding function f(x) for support/query examples, attention mechanism comparing query to support, weighted sum over support labels for prediction. Full Context Embeddings: Support set embedding uses bi-LSTM to read all support examples - embedding depends on context of other examples. Attention: Softmax attention with cosine similarity between query and support embeddings. Training: Episodic training on many N-way K-shot tasks sampled from training data, mimics test conditions. Comparison to Prototypical Networks: Matching uses attention (learnable), Prototypical uses mean (fixed). Matching more flexible, Prototypical simpler. Contribution: Introduced episodic training paradigm for few-shot learning, showed importance of test-time setup in training. Legacy: Influential paper establishing few-shot learning methodology, even if other methods now preferred.