Active Learning¶
---
title: Active Learning
---
flowchart LR
1[Label<br/>small sample] -->
2[Train<br/>model] -->
3[Predict<br/>unlabelled] -->
4["Query Selection<br/>(Rank poor predictions)"] -->
1 Query Selection Strategies¶
| Type | Strategy | Measure | Advantage |
|---|---|---|---|
| Heuristic | Uncertainty Sampling | Least confident Smallest margin Highest entropy | |
| QBC (Query By Committee) | Vote entropy of enseble models | ||
| Expected Model Change | Gradient of loss function wrt parameters | ||
| Expected Error Reduction | |||
| Variance Reduction | |||
| Density Weighted Methods | Average similarity to entire unlabelled pool of examples | Outliers ignored | |
| Approximate Posterior | Monte-Carlo Dropout | ||
| Stochastic Weight Modelling |