Confidence-Based Active Learning
summary
Presents a reduction from agnostic active learning to a confidence-rated predictor with nonzero error guarantee and gives an LP to construct that predictor (per-example alpha/beta abstain probabilities). Shows consistency and a label-complexity bound that replaces the disagreement coefficient by a smaller quantity sigma, recovering margin-based rates for log-concave / spherical data and improving over disagreement-based bounds; algorithm is asymptotically better but not yet computationally practical.
AI-written summary · about
chapters (18)
- 0:00Start
- 1:44Challenge: Acquiring Labeled Data
- 3:13Why Active Learning Helps?
- 6:03Agnostic Active Learning: Challenge
- 7:03Disagreement-based Active Learning
- 7:31Margin-based Active Learning
- 8:20Confidence-Rated Predictor (CRP)
- 9:10CRP with Guaranteed Error EW10
- 10:15Confidence-Based Active Learning
- 10:51Definition: Confidence Sets for c
- 11:35How to construct confidence sets?
- 11:58Active Learning via CRP
- 17:31Algorithm Outline
- 17:41Consistency Guarantees
- 18:30Label Complexity: Agnostic Case Parameters: Target error e VC dimension d Error of best cinc
- 21:53Linear Classification on Logconcave Distribution on Sphere
- 22:45Conclusions
- 23:41Open Question
related talks
43mBeyond Worst Case: When Complex Feedback Can Improve the Label Complexity of Active Learning
01h 31mInteractive Learning of Classifiers and Other Structures
37mEfficient Algorithms for Reliable Machine Learning
13mDistribution-Free, Risk-Controlling Prediction Sets
47mA Multi-Group Perspective on Fairness
01h 00mKnow When You Know: Handling Adversarial Data by Abstaining
33mOnline Multivalid Learning: Means, Moments, and Prediction Intervals
11mLinear Predictors in Reinforcement Learning: ...