Abstract
We identify structural properties under which a
convex optimization over the simplex can be massively
parallelized via map-reduce operations using
the Frank-Wolfe (FW) algorithm. A broad class
of problems, e.g., Convex Approximation, Experimental
Designs, and Adaboost, can be tackled this
way. We implement FW over Apache Spark, and
solve problems with 20 million variables using 350
cores in 79 minutes; the same operation takes 165
hours when executed serially.
Keywords
- Machine Learning: Data Mining
- Machine Learning: Machine Learning
- Computer Vision: Big Data and Large Scale Methods
- Machine Learning Applications: Big data ; Scalability
