Class meeting for 10-605 in Fall 2016 Overview

This is one of the class meetings on the schedule for the course Machine Learning with Large Datasets 10-605 in Fall 2016.

Slides

• Slides in Powerpoint
• Slides in PDF

Homework

• Before the next class: watch My overview lecture from 10-601 (lecture 1, and a little of lecture 2) if you need it.

• Today's quiz: [1]

Readings for the Class

• The Unreasonable Effectiveness of Data - Halevy, Pereira, Norvig

Also discussed

• William W. Cohen (1993): Efficient pruning methods for separate-and-conquer rule learning systems in IJCAI 1993: 988-994
• William W. Cohen (1995): Fast effective rule induction in ICML 1995: 115-123.
• Scaling to very very large corpora for natural language disambiguation, Banko & Brill, ACL 2001

Things to remember

• Why use big data?
  • Simple learning methods with large data sets can outperform complex learners with smaller datasets
  • The ordering of learning methods, best-to-worst, can be different for small datasets than from large datasets
  • The best way to improve performance for a learning system is often to collect more data
  • Large datasets often imply large classifiers

• Asymptotic analysis
  • It measures number of operations as function of problem size
  • Different operations (eg disk seeking, scanning, memory access) can have very very different costs
  • Disk access is cheapest when you scan sequentially