inductive learning

**Inductive learning** in the few-shot learning context refers to methods that classify each query example **independently**, using only the information from the labeled support set without considering other query examples. It builds a generalizable classification rule from the support set that can be applied to any new individual input. **How Inductive Few-Shot Learning Works** - **Step 1**: Receive the labeled support set (K examples per class). - **Step 2**: Build a classifier or decision rule from the support set alone. - **Step 3**: Apply this rule to each query example **independently** — the prediction for one query doesn't depend on any other query. **Inductive Few-Shot Methods** - **Prototypical Networks**: Compute class prototypes as **mean embeddings** of support examples. Classify each query by its distance to the nearest prototype. Each query is processed independently against the same prototypes. - **MAML**: Perform gradient-based adaptation on the support set to specialize model parameters, then apply the adapted model to each query independently. - **Matching Networks**: Weight support examples by similarity to each query using attention — but each query's classification depends only on its own similarities to support examples. - **Relation Networks**: Concatenate each query with each class prototype and pass through a learned relation module — independent per query. - **Simple Baselines**: Freeze pre-trained features, train a linear classifier or nearest-centroid classifier on support set embeddings. **Advantages of Inductive Approach** - **Streaming Compatible**: Works when query examples arrive **one at a time** — no need to batch queries. Essential for real-time applications. - **Consistent Predictions**: The prediction for a given query is **deterministic** — it doesn't change based on what other queries happen to be in the batch. - **No Distribution Assumptions**: Doesn't assume query examples cover all classes or follow any particular distribution. - **Simpler Implementation**: No iterative optimization or graph construction at test time. - **Lower Computational Cost**: Process each query in O(NK) time rather than O(N(K+Q)) for transductive methods. **Disadvantages vs. Transductive** - **Lower Accuracy**: Typically 2–5% lower than transductive methods on standard benchmarks because it ignores useful distributional information in the query batch. - **No Self-Correction**: Cannot use high-confidence predictions on some queries to improve uncertain predictions on others. - **Wasted Information**: The query batch often contains informative structure (clusters, density patterns) that inductive methods simply ignore. **When to Use Inductive** - **Real-Time Systems**: Predictions needed immediately as examples arrive — cannot wait for a full batch. - **Single Queries**: Only one test example available at a time (e.g., classifying individual images in a stream). - **Consistency Required**: Prediction for example X must not change depending on what else is in the test batch. - **Deployed Systems**: Production environments where simplicity and predictability are valued over marginal accuracy gains. Inductive learning is the **default approach** in most practical few-shot deployments — it trades a small accuracy penalty for simplicity, consistency, and compatibility with real-time and streaming applications.