Please send me an E-mail if you plan to attend Moed B.
Final Grades are published
Final details about the exam
|The exam will be held at Taub 5, from 9 to 12 on Friday, 30/1.|
You may bring lecture notes or whatever printed material you want.
|The tutorial will be held in room 201.|
Exam preparation tutorial
|It seems there is no time which suits everyone.|
I will hold the tutorial at Monday, 3-5 PM.
I will be in my office between 2-3 PM to answer any questions (The tutorial will be dedicated to solving exercises. I might do a very short review of SVM, but that's it)
More exercises from the booklet.
|Ex 3: Q1, 2, 3, 4 (Understand 4.2, no need to solve), 5, 9, 10, 12 (Note that the matrix doesn't have to be symmetric. understand the solution to 12.2 - No need to solve), 14 (understand solution. No need to solve), 15 (Note that in 15.3, it's actually true not only for the stored vectors but their negation as well).|
Note that many of these exercises are redundant.
Also, there is nothing about SVM or SVR in these exercises. I suggest you review the material on SVM (remember how you get the dual problem from the lagrangian) and understand the idea behind the kernel trick.
A note about VC dimension in the exercise booklet
|Some of the questions and answers in the exercise booklet which refer to VC dimension use a slightly different definition.|
The VC dimension is higher then the VC dimension you were taught by 1 according to that definition.
Next tutorial will be dedicated to exam preperation
|Also, I will be at my office between 15:30 and 16:30 instead of the usual hour.|
There will be another exam preparation tutorial (A longer one). see:
(Note that Tuesday might not be possible for me. I'm not sure yet)
If you'd like me to solve something specific, please ask by e-mail.
I will also list which exercises from the booklet are relevant soon.
More material has been uploaded
|Also, apparently, there were some exam solution files under "material" and they were set as "invisible".|
Grades updated & Announcement
|The grades for HW 3 have been updated and made visible (I forgot about the whole visibility thing...)|
Also, the final exam will be 3 hours, instead of 2 (as I might have said at some point)
|I have updated the grades.|
To the student who's ID ends in 58: I only got an answer to the first question (and barely that).
So I didn't give a grade yet.
Some things about backpropagation & the exercise
|In my rush to get to the exercise before the tutorial ended I forgot to say a few things about the backpropagation algorithm:|
1. When the algorithm converges, it converges to a local minima (addressing this is one of the possible ways to improve the algorithm)
2. When updating the weights, please use the function 1/2 * (t-y)^2, and when calculating the total error, please use (1/2n) * sum(t_i - y_i)^2
(just to make it more comfortable for me. If you did it differently and you'll need to re-run your code, you don't have to)
3. You need 1003 samples, not 1004. My bad.
New assignment has been published
|Good luck and happy holidays.|
Notes about SVM in HW2
|Please note that while SVM light does not use average error, other SVM implementations might (the sum of the error is divided by the number of training points)|
This might affect the values of C you might want to try (For example, C=400 might be too much, but when the error is divided by 200, 400 might not be high enough for C)
Added lecture notes
|Apologies for the delay, I am very forgetful :P|
HW1 has been graded & returned
|Returned according to the last 2 digits of the ID.|
|Submit by 10/12, please :P.|
Clarification on something in HW1
|If you're confused about me asking for both the number of epoches and the number of times we went over all the vectors, it's ok.|
No need for number of vector checks. Most people implement it in a way that makes the number of vector checks be (the number of vectors)*(number of epoches)
I just copied and pasted the assignment used by my predecessors (for the most part)
|I've uploaded my notes on lectures 2 and 3. Note that there may be some minor errors.|
I will scan more of my lecture notes next Thursday, and will probably upload them sometime that weekend.
Some clarifications about HW #1
|I've been asked some good questions, so here are some good answers:|
1. For Q2, you may assume the data is linearly separable.
2. You may program in any language you choose.
You're welcome to ask more questions by mail if you're not sure about something.
Also, next Wednesday, I will not be at the Technion during office hours. You are welcome to ask questions by e-mail.
If you'd like to meet at another time, send me an e-mail and we can try to set something up.
First assignment is published
|You should be done with the Perceptron by this Wednesday. You may, of course, start working on the assignment before that.|
Please submit the assignment by 19/11.
The website is up
|Apologies for taking so long with the website.|
The first tutorial will be held today at 16:30 in Taub 8.