Timeline

Wednesday, November 22

Abstract due (two pages).

Sunday, December 31

Project writeup is due. Up to 8 pages, in NIPS format. Look for "style files for creating camera-ready papers"; we recommend using LaTeX, but they also provide a template for Word.
No extensions!

Project format

The purpose of the project is to focus on a particular topic and explore it in relative depth. There are three major formats:

1. A survey.

   A detailed survey of previous work (as reflected in literature) on a clearly defined topic. This should focus on a relatively advanced topic, and therefore probably deal mostly with specialized publications (journal and conference papers) rather than textbook material. The survey should compare and contrast a number of approaches, and not just reiterate points made in each paper.

2. A novel application of machine learning.

   This should be a non-trivial application. E.g., just taking a classification problem, applying an arbitrarily chosen classifier and getting a result (which may be arbitrarily bad) is not an acceptable project.

3. Analysis of a machine learning algorithm or model. This can be empirical (well designed, in-depth study of the behavior of a model or algorithm which is not obvious) or theoretical, or both. This broad category includes non-trivial modifications of existing algorithms. E.g., you may identify a shortcoming in a standard technique and suggest a way to fix it, leading to improved behavior of that technique.

Projects in the last two categories could potentially lead to a novel contribution in machine learning and its applications, beyond this course.

Write-up instructions

The writeup should be written as a conference paper. This means that it should communicate the ideas, methods and (most important) conclusions concisely, but with enough detail for a reasonably knoledgeable reader to follow.

The project paper must include the following components.

• Introduction that clearly states the problem (what it is, and why it is important and/or difficult), and outlines on an intuitive level the solution proposed in the paper.
• Background, briefly describing related work and its relevance, emphasizing, if appropriate, similarity/differences with what you are doing to solve the problem.
• Technical description of the model/algorithm; this is the main "meat" of the paper.
• Experimental evaluation, if relevant. This should be concise but provide enough information for the results to be reproduced, with a reasonable effort.
• Conclusions. This is a very important part. In principle, a reader should be able to get a good idea what you did and what the take home message from your work is, by reading only the introduction and the conclusions sections.

The contents of a survey paper will of course be slightly different. The introduction in a survey should clearly state the topic, and outline the main conclusions. The

Data sets

Data repositories. These contain many data sets, mostly small ones, suitable for a large-scale comparative study of algorithms or for testing a novel algorithm or model.
• the Delve project at University of Toronto
• The UCI ML repository (has pointers to other databases as well)

Specialized data sets in particular areas.
• SPAM data sets: The SpamAssassin data set, the Enron data set.
• Handwritten digits: the MNIST data set
• Object categorization: the ETH-80 data set (simple 8-category object detection task)
• A more challenging object categorization task: Caltech 101 data set
• Document categorization: the Reuters-21578 data set
• Another document data set: the 20 Newsgroups data set