COGS 160 / 260 Semi-Supervised Learning

Latest changes to the site will be flagged here:     FINAL PAPERS DUE DEC 11 (SUNDAY)  -- approx 5-10 pages

SEE BELOW FOR UPDATED STYLE FILE IF YOU WERE HAVING TROUBLE WITH THE OLD ONE  (IGNORE IT IF YOU HAD NO PROBLEM)

                                                                              
                                                                                      
DEADLINES by Tues 11/29  email desa a group email (cc all group members) a plain text description of your project
(what animals/features you will use, what algorithms you will compare) and your current status (and any problems you have)

               by Thurs 12/1  email desa a pdf of your paper skeleton with motivation and methods filled in.
                    NIPS template style files are available here http://nips.cc/PaperInformation/StyleFiles
                    USE the following file instead of the .sty file found at that site to get your name to show up
                         nips11submit_e.sty
                    Latex is available on icogsci1-200 

              Final papers will be due  12/11.

                                Classes in Week 10 will meet in the computer room (CSB 115) and work on projects

Cognitive Science Building 003

Tues/Thurs 12:30pm – 2:00pm

Outline:

A common topic in machine learning is supervised learning (COGS118A): we have samples of data and a function associated with these samples (for example, images of objects and the category they belong to), and the goal is to approximate the function so that we can make predictions on future unlabelled datapoints. Unsupervised Learning (COGS118B) covers several techniques for learning from data with no labels, including trying to infer categories automatically (clustering) or creating a simpler set of variables that explain the observed data (dimensionality reduction, factor analysis, etc). 

This course is about boundary cases between the two, where we have a supervised task and something else:
  • additional unlabeled examples
  • another view of the data: for example a radically different set of sensors (sound and video for example)
  • access to the samples we wish to obtain predictions on, so instead of estimating a function we only need to estimate the value of the function at these points
  • an assumption the data lie on a nonlinear but smooth subspace
  • other tasks we can use to import inductive bias
This covers many techniques such as manifold learning, transduction, semi-supervised learning, multi-view learning, co-training, manifold regularization, spectral clustering, canonical correlation analysis. Our view is not exhaustive, but restricted to a subset of methods which share similar conceptual / mathematical roots. 

The goal of this course is to prepare students to conduct independent research in these areas. The course will be run as a seminar, reading some standout papers from the literature. Grading will largely be based on completion of a course project.

 Students are expected to have already taken an introductory machine learning course (e.g. one of  COGS 118A, 118B, CSE 150, 151, 152, ECE 175A, 175B or similar course).