Online learning, Vowpal Wabbit and Hadoop #bbuzz
summary
Claim: Vowpal Wabbit runs distributed online learning on Hadoop and speeds training; implemented by combining Vowpal Wabbit mappers with an all-reduce spanning tree over Hadoop, using EMR/S3 for transient clusters. Proven on a 6 GB test (≈15–50M examples, ~52 billion hashed features): single‑machine online pass runs ~26 minutes, parallel Hadoop run ~6 minutes. Caveats: Hadoop environment variables, non-splittable gz workaround, bootstrap/install and speculative‑task handling require extra engineering.
AI-written summary · about
chapters (20)
- 0:00Intro
- 2:44Linear regression
- 3:22Loss function
- 4:04Batch learning
- 5:38Parameter space
- 6:59Not fit in memory
- 8:29Learning
- 9:24Complexity
- 11:11Online vs Batch
- 16:15Summary
- 16:53Why paralyze
- 20:11All Reduce
- 24:37Spanning Tree
- 25:17Amazon Web Services
- 27:06Making it work
- 28:29Number of splits
- 29:21RAM allocation
- 31:13Benchmark
- 32:17Conclusion
- 33:24Coming soon
related talks
42mDown the Wabbit Hole
24mIntroduction to Hadoop
34mBig Data, Small Code ... #bbuzz
01h 10mGetting started with H2O on Python
17mMachine Learning at the Speed of Light with Spark
20mHow the cloud changed the way I do analytics and research
12mthe Rapidly Changing Hadoop Ecosystem
47mSolving Real & Big (Data) Problems Using Hadoop