Pattern Recognition at CSU East Bay, Winter 2009

syllabus

• course description: 6835: Pattern Recognition
• ask students who were in class - rest after rebootingcomputing summit

Notes

• http://www.nytimes.com/2009/01/07/technology/business-computing/07program.html?partner=rss
• PatternRecognitionVocab
• StatisticsReview
• Statistical programming with R from IBM developer's works
• PatternRecognitionHighLevelViews
• R package slides now on RisCool page

Labs

• how this kind of labs have developed in other classes:
  • WebDevLabsWinter08 - from back on Blackboard, this is a set of labs that are somewhat 'packaged' for other faculty, administrators, etc., to look at.
  • Labs from WebDevFall08 - 'in situ', not processed or modified at all, but you have to click to see each lab, etc.
• things we'd like to do differently
  • we'd like to get a much richer narrative. NandiniPremmanisakul asked whether investigators should be keeping a journal while doing a lab, and that is the perfect metaphor for the level of richness we want, at least for the moment.
  • we're trying to sort out some balance between prep work and involving the classes in the design of the labs
• our labs
  • SelfevalExampleTask
    • SelfevalExampleRoughDrafts
    • self eval example task final versions can just be linked to your home pages
  • SomLab
    • SomLabRoughDrafts
  • SvMLabDuh
    • SvMLabDuhRoughDrafts

What you should be doing

• developing your ability to scaffold (ScaffoldingInInformationScience). A good idea would be to add a thread to the narrative for your current lab that explicitly looks at your scaffolding process. What is particularly interesting about that approach is that it will strengthen two meta-cognitive skills: scaffolding, and reflection.
• SomLab, see above in Labs
• in class on Tuesday, practice (group) peer evaluations, see also LabPeerEvaluationTemplate, and PeerEvaluation?
• PatternRecognitionHighLevelViews
• working on R. See RisCool for info
• SelfevalExampleTask - work on for Tuesday
  • SelfevalExampleRoughDrafts

books to consider

• stack of books at first class, info after summit.

tools

areas of interest

Self-Organizing Maps

• Samuel Kaski's doctoral thesis on Data Exploration Using Self-Organizing Maps, see especially Chapters 2 (Methods for exploratory data analysis) and 3 (Stages of SOM-based exploratory data analysis).
• SOM in the wild, or an example of how SOM are actually used in a real project: Unobtrusive User Modeling for Adaptive Hypermedia (preprint)
• Mat Buckland's SOM tutorial
• SOM Bibliography from Helsinki University of Technology
• multiple contributed SOM packages in R, as well as multiple packages that include SOM
  • package list:
    • kohonen - a fairly modern implementation, as well as fairly complete (although missing some crucial elements, like sammon maps.) Includes a nice adaptation to using SOMs for classification.
    • wccsom - builds upon kohonen, adding weighted cross-correlation. Very nice, indeed.
    • som - a bare bones SOM implementation, has quantization error, but no Umatrix, no component maps.
    • class - contains batchSOM implementation (from 1995!).
  • reflective question: how do you decide which package to use? Some suggested criteria:
    1. package quality
       • reflective question: how do you assess the quality of a package? Some suggested criteria:
         1. does it include unit tests?
         2. how comprehensive is the documentation?
         3. ... (what else)?
    2. what is the goal of your overall project? does your goal match the goal of the package?
    3. ... (what else)?

Support Vector Machines

• Support Vector Machines for Classification and Regression - by Steve Gunn, is a more general discussion of SVMs, focusing on use and how they fit in to the bigger picture.
• A Tutorial on Support Vector Machines for Pattern Recognition - by Christopher J.C. Burges, has lots more detailed info on the different kernels and how you decide which to use. It also has excellent figures illustrating the optimal separating hyperplane for various types of kernels.
• lots of useful info at Support Vector Machines dot org, including a lovely visual explanation of the optimum hyperplane
• R resources for SVMs are listed in the Machine Learning task view at cran
  • More on e1071 (recommended, as it has all sorts of cool stuff)
    • Here's a nice tutorial from Hannes Planatscher and Janko Dietzsch
    • e1071 listing on crantastic. You can retrieve the documentation, search the r-help archives, etc.
    • you'll almost surely need to download and install libSVM separately. I built the C++ executables, rather than the Java ones, on my mac, as follows:
      1. I downloaded the .tar.gz archive, and moved it to my user account.
      2. I opened a terminal window and unpacked the archive from the command line
      3. I ran make.
      4. My R installation found libsvm without any further configuration. Nice.
      5. There's a Makefile.win for the windowz types out there...