The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Learning Active Learning with Ksenia Konyushkova - TWiML Talk #116
Mar 5, 2018
Ksenia Konyushkova, PhD student at EPFL focusing on active learning and annotation-efficiency. She discusses training models to choose which unlabeled examples most improve learning. Short simulations and real-data tests, strategies for hard checkerboard-like problems, smart bounding-box and segmentation workflows, and ways to speed human annotation with adaptive dialogs and batching.
AI Snips
Chapters
Transcript
Episode notes
Learned Active Learning Outperforms Fixed Heuristics
- Active learning can be learned from data by training a regressor to predict which unlabeled point will most reduce future error.
- Ksenia simulates many synthetic 2D datasets, records classifier state and point features, then trains a model to pick points that maximize error reduction.
Distance To Boundary Is Not Always The Best Criterion
- The learned selector uses many features beyond distance-to-boundary; classifier certainty and point density matter more.
- Ksenia found distance-to-boundary ranked fourth while classifier confidence was the top feature in her regressor analysis.
Adaptive Exploration Then Exploitation Beats Uncertainty Sampling
- A learned strategy adapts: when classifier is very uncertain it explores randomly, later it refines the boundary.
- In XOR/checkerboard problems uncertainty sampling can perform worse than random, but the learned policy discovers exploration-first behavior.