CS2750 - The TA Page

Course homepage

Check out the course homepage for vital information and assignments! Course homepage is the authoritative source, this one is just trying to be helpful. Check this page often though, because I post answers to questions mailed to me here so that everybody has equal information.

Announcements

Apr 03, 2004 Homework 8 is graded and ready for pickup.

image: decorative

Office hours

The official office hours are MW 10-11:30. Of course you are welcome to come almost any time I am in the office. To make sure I am available, you may want to send a mail or call 4-8832.

Assignments FAQ

Assignments are to be submitted according to these instructions via FTP.

Assignment 6

There is a glitch in the code given out of logistic_nn.m, towards the end of the file, computation of train/test errors. It's a copy/paste artifact and I'm sure you can fix it easily. It has been fixed in the code available for download.

Assignment 4

Answer:
Yes you can! Actually, what Matlab operator A\b is doing is roughly inv(A)*b. This only works for a regular (square) matrix, of course and we have (typically) more rows than columns.
A generalization of that operation (linear eqn system solving) happens to be (X^TX)^{-1}X^Tb, which you should compare with formula (3.6) in HTF. Use identity (FG)^{-1} = G^{-1}F^{-1} to convince yourself that for a nonsingular matrix this is the same as X^{-1}.

Question:
Where does the magic of '\' come from? Can I really do linear regression on one line of code?

Assignment 2

	    b) The mean squared error of a predictor is just a random
    variable, okay? The point here is to look at these two predictors
    and estimate the error each one is making. So you have 50 values
    of squared error, so you can compute the mean squared error, as
    well as the variance. That's all you need to plug into the formula.
    Just think of the "squared error" as a new variable.

    c) Hint: In this case you look at \mu_{diff} and quantify the
    certainty ("probabilistic support") that MSE_1 > MSE_2 (or
    vice versa) in terms of confidence intervals.

Tomas

On Tue, Jan 27, 2004 at 09:04:03PM -0500, ... wrote
> Dear Tomas,
>   Thank you for your last email - I understand how norminv works now.
> However, I really don't understand what the second and third questions
> of the homework are asking me to do. So you don't have to look it up,
> the questions are:
>   (b) Calculate & report confidence intervals for p = 0.95 for the mean
> squared errors of both predictors (Hint: use tinv or norminv).
>   (c) Which predictor appears to be better in the squared error sense.
> What is the probabilistic support for this hypothesis. (Hint: use tcdf
> or normcdf).

>I know what confidence intervals are and what mean squared errors are, but
>I don't know what confidence intervals _for_ mean squared errors are. And
>thus, I don't even know where to start with question (c).

Assignment 1

Answer:
I assume right now you are using the text-based Matlab version. Matlab displays the plots in a separate window. Make sure you have an X-server running at your machine. If you don't have one, you can have it installed by the tech staff in the department. In the command prompt, set the DISPLAY variable to <your IP>:0. Then run Matlab: 
	$ export DISPLAY=10.10.10.10:0
	$ matlab &
Replace 10.10.10.10 with your actual IP. (you can figure it out running 
ipconfig
from the command prompt of your Windows machine. ) You should get a graphical presentation of Matlab in a while. If you experience an inexplicable failure, check if there is a firewall in the way - it happened already.

Question:

	> Sent: Friday, January 16, 2004 11:03 AM
	> Subject: Matlab...
	> 
	> 
	> I've been trying to use matlab remotely from my account in my 
	> laptop at home(I have Windows XP, and I'm using fsecure to 
	> connect to unixs...)... I've tried the examples of the 
	> tutorial, and it's working fine ;) However, I can't display 
	> any graphs... no plots... is there any trick to do it? Beatriz