- Find a Partner. The topic talks will be done in groups of
two (or three, in special cases). That means that you'll have to find
another person to work with, and agree on a common topic. To make it
as easy as possible, we will set up a list of "personal ads," listing the students'
interests and backgrounds. Look through the list and identify people
who might share similar interests, then send them email and arrange to
meet with them so that you can come up with a topic.
- Find a topic. You can use the
list of potential topics as a starting
point. As soon as you have decided on the
general area (which can be fairly broad at this point), email that
info to Risto. Together Risto
and you will then focus it and come up with an initial reading
list. After the topics are set, Risto will put together a schedule of
talks roughly according to topic (see class schedule).
- Do the literature survey. Go over the reading material,
reading selectively, and follow the references. You might also want to
check the
tips on how to do literature surveys in neural nets.
- Assign the reading.
Two weeks before the talk meet with me after class; we'll decide on the
required and optional reading assignments for your talk and the general
outline of the talk. In general, one required paper (or two if they are
short) and 1-4 optional ones.
Email the references for both the required and the optional papers in
html format to Risto;
he'll put them in the reading
list. Make xerox copies of the required paper for everyone in
the class (including one for Risto), and bring them to the class one
week before your scheduled talk so we can distribute them.
The optional papers are distributed electronically through the reading list. If they are on the web,
please include the URL in the html reference you email to Risto (see
above); if they are available on paper only, either bring Risto a
paper copy or scan the article and send the pdf to Risto.
- Put together the talk. Put together a 50 min talk (about
20-40 slides total for the whole 50 mins). In general, you should (1)
give an intro to the topic, (2) present an overview of previous work,
(3) explain a few most interesting approaches in detail, (4) propose
your own ideas (your chance to get feedback for your project). Use a
lot of figures on your slides (almost every slide should have a
figure). Do not put too much on one slide (no full paragraphs). (Suggestions about presentations in general.)
Schedule a meeting with the TA through
email a few days before the talk. You should have your slides ready
on paper (printed 4-up, i.e. four slides per page) at that time, so
that you can go over the talk with the TA, and you should do it early
enough so that you can incorporate the TAs suggestions into your
talk. The idea is that the TA can give you feedback on how to make the
talk more accessible and useful for the wider audience.
- Grade the questions. Before the talk you will have a chance
to look at the questions and get an idea what the discussion is going
to be like, but it is not a test and you are not expected to be able
to answer all the questions. After your talk, you will grade the
questions (pass/fail). The idea is that your are in the best
position to judge them and benefit from them the most. You can also
provide feedback to the people who ask them, or to the whole cs394n
alias.
The questions should provide some new perspective or insight into the
material. Questions that concern trivial details (like "What was the
learning rate?" or have no clue ("I didn't understand how they trained
the network, please explain") should receive an F. Append your grades
to the list of students in the class,
and send the grades in email to the TA. Note: it is not sufficient to say
"everyone passed" because some people may not have turned in any
questions.