Reading: Chapter 4 of "Machine Learning" by Tom M. Mitchell. Note, Sections 4.5.3 and 4.8 are optional. However, section 4.8 may provide ideas for the Part II of this assignment.

Your assignment: Follow the instructions for Mitchell and Blum's "Neural Networks and face images" assignment. The first level of detail for this project is described in section 4.7 of Mitchell's text. The remaining details for the NN implementation and assignment are in hw96.ps, attached if you got this handout in class. You are provided with a NN implementation in C, which you will have to modify a little. Note that this assignment description says "Homework 5", but this is actually your third assignment, and also note the above due date, in contrast to that in the description.

Like decision trees, this assignment has two parts, the second of which is open-ended. For this assignment, we strongly recommend that you team up for the second part. For the second part you will optionally make a brief presentation in class, and will hand in a write-up.

Part I:

Changes from the assignment description file:

• Files are available on ACiS from the TA's account at:
  • The implementation C code: ~dke3/cs4771/code/
  • The (small) pictures to train on: ~dke3/cs4771/data/faces_4/
  • The lists of file names to train on: ~dke3/cs4771/data/trainset/
• Remote students with no AcIS account should arrange with Dave to get the code and data.

For Part I you are to hand in answers to the questions listed among the 13 steps described in the assignment.

Part II:

For the open-ended portion of the assignment, we strongly recommend you team up with one or two other people. This way, you can build on each other's ideas, share the implementation and write-up responsibilities, and have more fun. For this portion, you hand in one write-up for your group.

Since you do not have to write the implementation from scratch, emphasis is weighted more on Part II of this assignment than it was for the decision tree assignment. However, do be careful again to keep your experiments very reasonable in terms of effort -- there are still two more projects after this one. If you have "big ideas", they may best wait until the final project, which is, in principle, completely open-ended (i.e., can use NNs, decision trees, GAs, anything).

Please choose something that you find interesting. Your choice should be motivated, not just random. You should be able to explain why your experiment was an interesting thing to try out.

In addition to the suggestions listed in the assignment description file, other possibilities include:

• Apply the package to your own data, i.e., image data or otherwise. For example, cameras in the SRL lab will allow you to take your own pictures of faces or other objects. However, you will need help from someone with a computer science account. If you collect this data, you may want to continue using it for your final project. fyi, a CRF person said: The software is in /opt/SUNWits and /opt/SUNWrtvc. I think you can use "SunVideo", "rtvc_display", "rtvc_capture_movie"
• One student at CMU reportedly created a "sunglass learner/generator" that placed sunglasses on a sunglasses-less face.
• NN for hand signal recognition

Deliverables for Part II:

• a written description of the experiments you conducted, 4 page max. This should describe the details of your modification, the motivation ("why?"), and the resulting change of performance. As supporting reference material for this, also append:
  • hard copy of additional or modified code (well documented). Please organize your code well.
  • hard copy of sample runs (also well documented)
• 2 slides, one on what you did, and one on the results. I or One or more members of your group will "flash" these slides for the class and describe your work. Please e-mail me if you are interested in doing this (we do not have time for everyone to present -- I will do quick overviews of the remaining ones myself) -- it will be fun to present. To presenters: You will only have a few minutes to do this, so you need to focus on the important aspects of what you've done. If this makes you feel nervous, that is normal, and you are welcome to speak with us about it. Keep in mind that the purpose of this is to share ideas in a fun way, and you are not expected to "perform" well, but simply to convey some basic information to the class.
• Please email your writeup, any supporting materials (slides, etc) and any modified to the TA, Dave Evans at devans@cs.columbia.edu. The submission can be in any reasonable format, such as a gzipped tar archive, a windows zip attachment, a Macintosh stuffit archive, etc. I consider submissions in Disk Disintegrator Deluxe (Apple ][) or obscure Amiga formats to be unreasonable. Similarly for the writeup, MS Word, LaTex, HTML, and so on are reasonable. Wordstar is not.
• (a brief demo may also be requested later)

email: evs at cs dot columbia dot edu