Syllabus for Machine Learning with Large Datasets 10-605 in Spring 2013

This is the syllabus for Machine Learning with Large Datasets 10-605 in Spring 2013.

January

• Mon Jan 14. Overview of course, cost of various operations, asymptotic analysis.
• Wed Jan 16. Review of probabilities.
• Mon Jan 21. no class - Martin Luther King Day
• Wed Jan 23. Streaming algorithms and Naive Bayes; The stream-and-sort design pattern; Naive Bayes for large feature sets.
  • New Assignment: streaming Naive Bayes 1 (with feature counts in memory). PDF Handout
• Mon Jan 28. Messages and records 1; Phrase finding.
  • Assignment due: streaming Naive Bayes 1 (with feature counts in memory).
  • New Assignment: streaming Naive Bayes 2 (with feature counts on disk) with stream-and-sort. PDF Handout
• Wed Jan 30. More on streaming algorithms: Rocchio, and theory of on-line learning

February

• Mon Feb 4. More on streaming algorithms: parallelized voted perceptrons.
  • Assignment due: streaming Naive Bayes 2 (with feature counts on disk) with stream-and-sort
  • New Assignment: phrase finding with stream-and-sort. PDF Handout
• Wed Feb 6. Map-reduce and Hadoop 1.
• Mon Feb 11. Map-reduce and Hadoop 2.
• Wed Feb 13. Hadoop helpers and Scalable SGD
  • Assignment due: phrase finding with stream-and-sort
  • New Assignments: Naive Bayes with Hadoop & Phrase-finding with Hadoop. PDF Handout
• Mon Feb 18. Scalable SGD and Hash Kernels
• Wed Feb 20. Guest lecture: Chris Dyer. Scalable feature selection with Map-Reduce
  • Streaming run on Hadoop of Naive Bayes due - checkpoint
• Mon Feb 25. Background on randomized algorithms; Graph computations 1.
• Wed Feb 27. Guest Lecture: Aapo Kyrola - GraphLab and GraphChi
  • Hadoop assignment (Naive Bayes) due

March

• Mon Mar 4. Learning on graphs 2.
• Wed Mar 6. Guest lecture: John Wong (Google): Machine Learning with Large Datasets in Google Shopping"
  • Hadoop assignment (phrase-finding) due
  • New Assignment: memory-efficient SGD PDF writeup
  • New assignment: initial project proposals. PDF writeup
• Mon Mar 11. no class - spring break.
• Wed Mar 13. no class - spring break.
• Mon Mar 18. Subsampling a graph with RWR
• Wed Mar 20. Semi-supervised learning via label propagation on graphs
  • Assignment due: initial mini-project proposals.
  • Assignment due: memory-efficient SGD
  • New Assignment: Subsampling and visualizing a graph. PDF writeup
• Mon Mar 25. Label propagation 2: Unsupervised label propagation, label propagation as optimization, bipartite graphs
• Wed Mar 27. Understanding spectral clustering techniques.
  • Assignment due: mini-project proposals (final version).

April and May

• Mon Apr 1. Speeding up LDA-like models: sparse sampling and parallelization
• Wed Apr 3. Speeding up LDA-like models: All-reduce and online LDA
  • Assignment due: Subsampling and visualizing a graph.
  • New Assignment: K-Means on MapReduce. PDF writeup
• Mon Apr 8. Fast KNN and similarity joins 1.
• Wed Apr 10. Fast KNN and similarity joins 2.
• Mon Apr 15. Scaling up decision tree learning
  • Project progress report due
• Wed Apr 17. Gradient boosting with trees, and SGD for matrix factorization
  • Assignment due: K-Means on MapReduce.
  • New Assignment: Multi-class image classification or scalable classification using a linear classifier. Both of these count as one assignment toward your six.
    • PDF writeup of image-classification assignment
    • PDF writeup of scalable classification
• Mon Apr 22. Guest lecture, Evangelos Papalexakis, on Scalable Tensor Methods.
• Wed Apr 24. Project reports.
• Mon Apr 29. Project reports.
• Wed May 1. Project reports.
  • Assignment due: Multi-class image classification or scalable classification.

May

• 9am, Tuesday, May 7. Project writeups due. Submit a paper to Blackbook in PDF in the ICML 2013 format (minimum 5 pp, up to 8pp double column), except, of course, do not submit it anonymously.
  • Note: this is extended from previous deadline of Fri May 3---but I can't give any further extensions! Your project report should discuss
    • The problem you're trying to solve, and why it's important and/or interesting.
    • Related work, especially any related work that you're building on.
    • The data that you're working with.
    • The methods that you're using (in some detail - even if these are off-the-shelf methods, I want to know that you understand them)
    • The experiments you did, the metrics you used to evaluate them, and the results.
    • What was learned from the experiments (the conclusions).
  • You should think of this as an exercise in writing a conference-style paper: so try and write in that style. (Of course, your work doesn't need to advance the state-of-the-art in machine learning, or be highly novel, but it should be well-described.)