Multiscale Methods for Machine Learning

Explore a 55-minute conference talk from the SIAM Conference on Applied Linear Algebra (LA21) focusing on multiscale methods for machine learning. Delve into the Prediction for Enormous and Correlated Output Spaces (PECOS) framework, presented by Inderjit S. Dhillon from the University of Texas at Austin. Learn how this innovative approach addresses the challenge of quickly finding relevant results from vast output spaces in modern applications. Discover the multi-scale machine learning techniques that build hierarchies over output spaces using unsupervised learning and combine predictions at different levels. Understand how PECOS leverages sparsity and efficient sparse matrix routines to achieve logarithmic scaling with output space size. Gain insights into weighted graph cuts, weighted kernel K-means, semantic indexing, and masked sparse chunk multiplication. Explore the experimental results and implications of this groundbreaking methodology for handling enormous and correlated output spaces in machine learning applications.

Introduction
Outline
Frameworks
Linear Methods
Multioutput prediction
Textbooks
Methodology
Challenges
Weighted Graph Cuts
Weighted Kernel KMeans
Semantic Indexing
Machine Learning Training
Beam Search
Masked sparse chunk multiplication
Sparse column format
Data structure
Cache efficiency
Experimental results
Conclusion
Questions
Closing remarks
Thank you
Attendance figures