Learning from Labeled and Unlabeled Data

Motivation: limited labeled data cost

Motivation for weak supervision (semi-supervised learning) is that obtaining labeled data is often expensive, time-consuming, or impractical because it may require skilled human effort (e.g., transcription) or costly physical experimentation (e.g., determining protein structure or detecting oil). As a result, large fully labeled training sets may be infeasible, while collecting unlabeled data is comparatively cheap. Weak supervision addresses this by training models using a small subset of human-labeled (or otherwise precisely labeled) examples together with a much larger set of unlabeled (or imprecisely labeled) examples. The goal is to improve predictive performance beyond what would be achieved by either (1) discarding the unlabeled data and training only on the limited labeled set (supervised learning) or (2) discarding the labels and relying only on unlabeled data (unsupervised learning).

0/2 modes complete

Semi-supervised learning paradigm overview

0/2 modes complete