For your project, you should investigate some interesting aspect of machine learning.  This should include

• a broad thorough literature review , overviewing this general topic;
• Ex1: techniques for learning motifs in DNA
• Ex2: ways to cope with missing data
• a deeper discussion of some specific subtopic
• Ex1: using hidden markov models to learn probabilistic motifs
• Ex2: statistically motivated ways to handle independently blocked attribute values
• an analysis of several systems for this task -- theoretical and/or empirical (preferably based on your implementation, or at least your runs on various data)
• Ex1: an empirical comparison of several gene-finding tools, on novel datasets
• Ex2: an empirical, and perhaps theoretical, analysis of several specific techniques, on novel data

Note that this investigation may begin by reading two or more recent, related papers from conferences/journals on artificial intelligence, where the papers are related by tackling the same problem but using different approaches; or by employing similar techniques to solve different problems; or as one is a follow-up work of the other; or they hold the opposite points of view over some problems; etc.
 

Evaluation Criteria.

• Deciding on some specific topic
• Reading the relevant literature
• Implementing some of the ideas and gathering experimental data
  and/or providing some new theoretical insights
• Writing an original 8-page paper
• critically overviewing the literature
• presenting your results
[Due about two weeks after the last day of class.]
There are 3 ways to hand-in your final project report:
• Slip the hard-copy under my door (Ath 359)
• If it is in *.pdf or *.ps, you may EMAIL the file to your coach
• If it is *.pdf, *.ps or *.html, you put a copy on a webpage, then EMAIL that URL to your coach.
Or you can do more than one of the above...
N.b., do NOT send me *.doc,... files!
See RequiredComponents
• Giving two presentations to the class (either during class or lab time; see schedule for dates, and "teams" for specific times for each team).

• The first, around week#9 [~10/Mar; ~15/Nov], will summarize the area of your project -- giving a "lay of the land"; see Guide .
  Each member of the audience will fill out this Feedback form

• The second presentation will summarize your results.  It will be given during the last week of the semester. (Again, see Guide .)
  Each member of the audience will fill out this Feedback form

Each presentation will be 15 minutes: 12 minute presentation and 3 minutes for questions (+ set-up)

• Apparent effort.
• Clarity. Does your paper demonstrate a good understanding and give a good analysis of the underlying challenge?  Is it well-written and well-organized?  Does it make good use of examples to illustrate the problem and the solutions?
• Originality.  Does your paper show independent thought and consideration, identify any unsolved problems and propose any initial solutions, or propose any new approaches to existing problems?

Rough Guidelines

Required Components of WriteUp.

• Clear statement of problem being addressed
  • clearly describes the underlying performance task, and distinguishing this from the learning task.

    It is often (?always?) useful to include one figure that explicitly shows the input and the output of the performance system, and another figure showing the input and the output of the learning system.

• Motivation -- i.e.,  Why is this problem interesting and challenging?
• Necessary background material, review of previous work and limitations of previous work.
• Clear statement of technical solutions used to solve the problem, how successful they were, and why they were successful.
• use clearly defined terms. It is fine to use intuititions to motivate an idea, but thereafter use include precise statements to state your claim(s), then support these claims with either theorems or meaningful empirical results.
• Identification of remaining problems and future research directions.
• in the form of a NIPS paper (8 pages maximum in NIPS format, including references; see NIPS Style)
  But DO include your name!

What was the problem you worked on,
why is it important,
what did you learn, and
why are these results important.

Try to make your report EASY to read.

In a nutshell, your write-up tells a story, in a clear fashion. Your paper should work to establish some explicitly-stated "falsifiable conclusion" (which should, of course, be related to learning...) Every section, paragraph, section, figure, table, ... should contribute to establishing this specific claim. Towards enforcing this, your first section should include an overview, outlining the contents of the paper. You may also want to begin each section with an overview, indicating both what will be included here, and also connecting this to the central theme. As an example, suppose you are claiming that algX works effectively at taskY (eg, algX=="Support Vector Machines", taskY=="detecting patterns in heart rhythms"). Here, it makes sense to describe algX, and perhaps its precursors, and to contrast algX with other related algorithms. (Note this contrast is typically in the form "algQ does BLAH; our algX differs by doing the subBLAH differently, our report proves that this is an improvement"; etc.) You should also discuss the effects of changing the settings for various parameters. Similarly, you should precisely define taskY, and perhaps contrast it with other related tasks. You should then provide evidence to establish the claim -- either empirical or theoretical.

If your report contrasts algQ with algR, you should explain why that is relevant. Or if it digresses to consider some taskW, again explain why this is included. (If you simply want to include such analysis -- perhaps to indicate that you had read an article -- you may include it in an appendix, possibly labeled as "not completely irrelevant asides" :-) )

Your report should contain precisely-defined terms; do not be afraid of using mathematical notation! Similarly, if you use comparisons, be sure to specify the details; eg, state "algX is an improvement over algZ", rather than just "algX is an improvement".

You should include simple illustrative examples! One that conveys the basic ideas, to help the reader understand the various points.

Be sure to re-read your report! Imagine this topic was new to you... would you understand the material presented? You may assume your reader knows only material presented in 466/551; if you use any other terms, be sure they are defined. You should also explain why you are including that term -- ie, how does it relate to the overall theme of the paper.
Don't make your reader guess at your meanings!

Be sure to label figures/tables. (Eg, if you write "10%", is this 10% error, or 10% accuracy?)

NOTES:

• REQUIREMENTS:
  Each report must include a simple specific example, providing the I/O showing how the output is related to the input specifying the desired/achived properties of the output illustrating the basic terms used.
  It must also include an outline of the paper (probably at end of section1), that specifies the goals of the research, and overview the paper.
• There is often a large number of parameters that can be adjusted; typically too many to exhaustively try every combination. Here, you should still explore the space. First, you can argue that some specific parameters appear largely irrelevant -- here, by considering a few specific setting for n-1 parameters, and for each such setting, varying the remaining parameter. If the result for each setting does not change, we can typically ignore that extra parameter. Second, you can in general look for correlations amoung the parameters. Etc.

Logistics

• I encourage people to work in teams of three or four; note that all Grads in a team will receive the same grade, as will all Undergrads.
• Your should ALSO hand in a 9th page, that quickly summarizes who did what for this project -- wrt conceptualizing the problem, implementing the algorithms, running experiments, proving theorems, writing the final document, preparing the presentations, etc etc etc. This can be a simple table (perhaps summarizing which team member did what percent of each -- in round figures!) or whatever is easiest for your team.

• Decide on topics, and teams, by around 5th week of term (~5/Feb; ~6/Oct).
  Each team must email to me a one-page, PLAIN-TEXT document that provides:
  • Title of your project
      + Proposed topic
  • Team members (also specify Undergrad/Grad, and major, as well as email address)
  • What you plan to do, or avoid doing
  • Papers you plan to read/critique
  • Code you plan to write (or download)
  • Experiments you plan to perform
  • Rough outline of who will do what; and estimate of time requirements
  • When your team can meet with me, for our bi-weekly meeting. (Suggest 3 times.)
    See my schedule for times to avoid.
• Bi-weekly meeting with me thenafter.
Issues for First Meeting [week of 11-15/Feb; 12-16/Oct]
• Make sure the task and evaluation criteria are both well-defined
• If the task involves data:
  Do you have it? How will you get it?
• Do you have the resources you need?
• General questions?
Issues for Second Meeting [week of 3-7/Mar; 26-30/Oct]
• Follow-up on "Issues for First Meeting":
• Well defined task, data, resources, ...
  
• Prepare for Presentation#1: Lay of the Land
• General questions ?
Issues for Third Meeting [week of 17-24/Mar; 9-13/Nov]
• Continue preparing for Presentation#1
• General question ?
Issues for Fourth Meeting [week of ~1/Apr; 23-27/Nov]
• Begin preparation for Presentation#2
• General question ?


• If your project is an empirical study, see below .
• Presentation#1 will be around Week#8.
• Presentation#2 will be during the last week of the class.
• The final write-up (discussing empirical results, etc etc etc) is due two weeks after the end of the course.

• the write-up [eg, *.ps, *.pdf, *.doc, or whatever]
• any other files you'd like me to have eg, data, algorithms, charts, ...
  Be sure each is labeled appropriately.

Many people are considering empirical studies -- eg, "application pulls". Here, the learning challenge is

how to use some "experiences" to improve "performance" on some "performance task".

The Learning Task

• a PERFORMANCE TASK [eg, playing poker, driving a car, ...]

Notice this is independent of learning -- eg, one could build a NON-LEARNING performance system that does this.
• a well-defined, objectively-measured PERFORMANCE CRITERIA, for evaluating a performance system

[eg, %ofhands won, or "average number of miles driven before accident".
Note: "kinda seems good" is not objective :-) ]
• the type of "EXPERIENCES" the learner will use, both the "instances" and how they are "labeled"

  [eg, the instances could be the poker hands played against a random opponents, or perhaps against a series of progressively cleverer opponents, or ... The "label" could be perhaps a numeric score indicating the quality of each action for a given hand (eg, a label of -1 for "raising" on 4h8d, or "0" for folding, etc) or it could be feedback after a SEQUENCE of actions (eg, this sequence of moves led to a lost game; or the plant was fined at time 7, or ...); or ... ]

• how you will evaluate the LEARNER

  Typically this will be in terms of a learning curve: after experiences S, learner L will produce the performance system L(S). We can then measure the true error rate for L(S) on new data. ... we can also consider, eg, how L improved the performance L(S) - L({}), or typically measure the EXPECTED value of L(S)-L({}) over a range of "training sets of size m=|S|".
  You should also include statistical tests to show whether your claims are statistically value. Here, be sure to explicitly state which test you ran, and the confidence --- as in "... this result is significant based on a paired t-test, p<, 0.05."

• the actual type of performance system (eg, decision tree, rule set, neural net, ...); and
• the actual learner involved (eg, reinforcement learning, consistency filtering, backpropagation, ...)

• suggest the actual experiments that will be run, typically comparing different learners for some learning task. (Eg, Neural Nets vs Decision Trees, or Reinforcement learning vs ILP, or ...)
• what literature you plan to read as well as logistic issues, etc etc etc.

Other comments

• Don't be afraid to use formulas!
• Think of how the material should be organized.
• Be sure to distinguish LEARNER from CLASSIFIER
• If you don't use some aspect, you don't have to discuss it!
• Abstract != overview
  The abstract should say what the results are!
• Think of your target audience... provide information that they need to know, and only information that they do not know.
• See Hints about preparing PowerPoints documents.