Parallel Active Learning
Talkat Parallel and Distributed Algorithms for Inference and Optimization (Simons Institute, Oct. 21–24, 2013) · 2013-10-21..10-24·
summary
A parallel active learning pipeline combines a local active sifter that scans stored examples and subsamples informative points with a global passive updater that retrains on those samples; implemented for RBF kernel SVMs on an expanded MNIST (≈8.1M warped images) and a single‑hidden‑layer neural net. Sequential active learning already cuts labelled data dramatically; adding parallelism yields near‑linear speedups up to ~100 nodes (simulation, 1,000 examples per class in the experiment), while neural nets show smaller gains.
AI-written summary · about
chapters (7)
- 0:00Motivation and context
- 1:54Implicit limitations
- 5:24Active learning approach
- 8:49Proposed parallel algorithm
- 12:42Complexity analysis
- 17:07Asynchronous challenges
- 20:28Experimental results
related talks
36mActive-Active Machine Learning Infrastructures For Training And Prediction
37mActive Teaching, Human Learning
35mScikit-Learn can do THAT?!
50mBuilding Continuous Learning Systems
34mDeep Learning at Scale
39mActive learning in the interactive python environment
19mAccelerating the Training of Large Language Models using Efficient Activation...
18mActive Learning for Neurosymbolic Program Synthesis