Numbers

Due: April 19, 2008 23:59:59 EST

Please have one of your group members send this assignment to your TA with the subject line "cs7641finalproject". Please include the names of everyone in your group in the body of the email message.

Why?

In some sense, the whole point of this class was to give you an opportunity to explore a machine learning research project. Given that, we take this assignment very seriously, as should you. Of course, if you haven't up to this point, it is already much too late, but you should try anyway.

What to turn in

We will define gt to be the gt account of one of your group members. It doesn't matter which one, as long as it is the same name you used last time.

You must email the TA a tar or zip file named gt.{zip,tar,tar.gz} that contains a single folder or directory named gt that in turn contains:

1. The Final Paper. It should be named gt.pdf. The file must be a pdf file and must be in the NIPS 2005 format. This includes the NIPS style, abstract, page count, references. In short you are answering the NIPS 2005 Call for Papers and turning in a conference paper for consideration.

   You should have already turned this in anyway. You will have received reviews by now and should incorporate any changes that make sense into your paper and your work. You should include a document called gt-errata.pdf that outlines the changes from your first submission and your final paper. It should be short.

2. The Final Presentation. It should be named gt-slides.{ppt,key{.zip},pdf}. You are strongly encouraged to turn in your final presentation in Powerpoint or Keynote. Barring that, please turn in a pdf. Remember: these are the final slides that you will be using for your talk. There will be no last-minute substitutions.

3. Any supporting material. You may want to keep a local copy of a movie file or run a demo or whatever. That's okay; put those files in your folder. Also, because this is your final chance to turn in everything from your project and convince us that you did work, you might want to include supporting files like code, extra graphs, your data, or whatever.

It is your responsiblity to make sure that the paper and presentation are readable. Your target machine is an Apple OS X Macbook Pro.

Helpful hints

• Practice. Please practice your talk several times before showing up. Try to anticipate questions. This is a professional talk after all. Let's not make it more painful than necessary.
• Time. You will be given only a limited time. Currently, that looks to be 20 minutes per talk. That seems like a lot, but that's barely enough time to say your name, much less describe the problem, the approach, and the conclusion. On the other hand, it is enough time to make it obvious that you haven't properly prepared. I highly recommend that only one of the group members actually speaks. I also recommend that you time your talk, and factor in a few questions.
• State your conclusion early. It is a common mistake to treat a talk like a mystery novel with a surprise ending. Given the time constraints, I recommend making your first slide an overall statement of a larger take-home message (e.g., "Statistical ML is ripe for application to human-computer and human-human interaction") perhaps with a picture or graph; your second slide a description of the problem your are tackling (e.g., "Modelling Human Interaction in Email"), again with a nice picture/example; your third slide the conclusion (e.g., "HMMs provide a decent model of motifs and flow of conversation in email but need too much data to extract person-specific information"); the middle slides your approach, experimental set up, and results; and your final slides a reiteration of your conclusion and possibly future work. In other words: tell me what you're going to tell me, tell me, then tell me what you told me. Skip the outline slide. Please.
• Know your audience. Your audience is a group of relatively bright listeners who have had at least a semester's worth of exposure to machine learning. There are many of them so you have to be mildly entertaining, but they aren't so many of them that they won't feel the urge to ask you clarifying questions. Assume they are smart but don't know the details of what you're talking about, so you have to give them enough background so that they can appreciate the brillance of your work, but not so much background that they feel like they're in first grade.
• Go read successful NIPS papers. This is not a proposal, it is a conference submission. Look at accepted NIPS papers, and note how they are laid out.
• Situate your work. We do related work sections not just to prove that we can read but because it helps us to place our work in the much larger body of research that is out there (oh, and to show that you've done due diligence). Make connections. You may not have time to do a detailed version of this in your talk, but you most certainly must in your paper (Note: "may not have time to do a detailed version" is not the same thing as "do not do any connections in the talk").
• Start yesterday. If you do not go through two or three iterations of your paper and one or two of your presentation you are either geniuses or idiots (I'll let you guess the prior on that). All of this requires time. Start right now.