Animashree Anandkumar, Electrical Engineering and CS Dept, UC Irvine @ MLconf SEA
summary
Tensor methods claim consistent parameter recovery for topic models, community/mixed‑membership models, HMMs and ICA by decomposing third-order moments instead of optimizing likelihood, and scale via simple SVD and matrix ops so they parallelize on CPU/GPU and cloud. Implementations run on Apache Reef and a spark-based implementation is in testing, with demo results on text, social graph, healthcare and mouse‑brain cell data and measured speedups of two or three orders of magnitude versus variational likelihood methods.
AI-written summary · about
chapters (12)
- 0:00Introduction to speaker
- 1:06Challenges of big data
- 2:08Tensors vs matrices
- 3:36Tensor decomposition methods
- 6:44Computational implementations
- 9:33Applications of tensors
- 11:58Latent variable modeling
- 14:45Consistent model learning
- 18:33Framework for latent models
- 22:01Practical demos
- 26:46Community and graph analysis
- 29:48Future directions
related talks
22mAnima Anandkumar, Electrical Engineering and CS Dept, UC Irvine
52mTensor Methods for Learning Latent Variable Models: Theory and Practice
56mTensor Decompositions for Learning Latent Variable Models I
16mAmeet Talwalkar, Assistant Professor of Computer Science, UCLA @ MLconf SF Part 2
16mAndrew Musselman, Committer & PMC Member, Apache Mahout at MLconf Seattle 2017
30mHanie Sedghi, Research Scientist, Allen Institute at MLconf Seattle 2017
42mAlex Smola, Professor in the Machine Learning Department, Carnegie Mellon University
25mAnusua Trivedi, Data Scientist at Texas Advanced Computing Center (TACC), UT Austin