Large-scale parallel implementations of SVMs

Posted in Science on August 17, 2008


Large-scale parallel implementations of SVMs

Lecture slides:

  • Related Work
  • The Problem: Regular, plain SVM
  • The Algorithm
  • Engineering
  • Engineering in practice
  • Vector-type optimized Kernels
  • Sorting the data set by labels
  • Multi-threading on multi-processor machines
  • Paralelization: Spread-Kernel: full data [2 nodes]
  • Paralelization: Spread-Kernel: full data [p nodes]
  • The network max( WorkingSet ) [p nodes]
  • Paralelization: Spread-Kernel: split data [2 nodes]
  • Paralelization: Spread-Kernel: split data [p nodes]
  • Reliable MULTICAST
  • NEC Cluster
  • Results: Speedup: theoretical model
  • Results: Speedup: Training Forest [522K samples]
  • Results: Speedup: Training MNIST [220K samples]
  • Results: Speedup: Training MNIST [500K samples]
  • Results: Speedup: Training MNIST [1M samples]
  • Results: Speedup: Training MNIST [2M samples]
  • Results: Speedup: Training MNIST [4M samples]
  • Summary
  • Software [availability to be determined]

Author: Igor Durdanovi?, Nec Laboratories America, Inc.

Watch Video

Tags: Science, Lectures, Computer Science, Machine Learning, VideoLectures.Net, Kernel Methods