CS 260: Machine Learning

Hi everyone,

     I will have an extra office hour tomorrow (Dec 11) 1:00-2:00pm in my office (EVI 282). Feel free to stop by. Thanks.

Best,

Quanquan

Hi all,

We will hold our final presentation on

Wednesday, December 12, 3:00pm - 6:00pm

MS 5200

Before you prepare the presentation, please read the following instruction carefully:

1.     Timing. You will be expected to deliver a polished and concise overview of your final project in the 12 minute time frame.  We will be very strict about timing and will stop you when your 12 minutes time is up even if you have not finished. You are recommended to prepare a 10-minute presentation, and leave the rest 2 minutes for questions and answers.

2.     Order. Please complete the survey https://doodle.com/poll/tyn6cq7at7k3i55i by Friday, Dec 7th, 11:59 PM to select your preferred slot. Instruction of the survey can be found here https://piazza.com/class/jme0w759wny7k?cid=214. Teams failed to complete the survey before the deadline will be randomly assigned a slot.

3.     Content. Your presentation should generally reflect the outcome and accomplishment of the project described in your proposal.

4.     Audience. Due to the time limit, when listening to other groups’ presentation, please avoid interrupting the speaker until they finish. You are welcome to ask questions in questions & answers section.

5.     Slides. Due to the time limit, all teams are required to use one single computer which is tied to the video projector of the classroom. Please submit all materials you want to display during your final presentation to CCLE. The deadline of submission is Dec 12nd, 2:00PM (1 hour before the final presentation). Any changes after the slide submission deadline will not be considered.

1.     We recommend that you use either PowerPoint, Keynote, Adobe Acrobat or browser-based applications as these tend to be available on most computers. To avoid layout issue during display, it is recommended to export your slides into print-ready file such as high-resolution Adobe PDF files.

2.     If animations or videos are necessary, please ensure that you include the corresponding files in your submission. We encourage you to submit videos using an MP4 codec such as DivX contained in an AVI. Also, please submit a README text file with each video specifying the exact codec used and a URL where the codec can be downloaded. Another option is you can upload your videos to websites like Youtube, and insert or link the online video to your slides.

3.     CCLE imposes a 100MB limit on the size of submission. Note that you can update the file by uploading a new one.

4.     If you have any special demand for technical support, please contact TA (xzbei@cs.ucla.edu) at least one day before the final presentation. 

Good luck!

#pin

Hi all,

The final presentation day & time will be :

Wednesday, December 12, 3:00pm - 6:00pm

We are going to use the doodle poll to set up the order of the final presentation. Please read the instruction carefully before you sign up:

1. First come, first choose.

2. One team can only submit once.

3. Once sign up, you are required to fulfill your name. You only need to fulfill the name of one of your members in your team.

4. You can pick only one slot. One slot can be chosen by at most one team.

5. Please avoid re-editing. 

6. Teams failed to submit the poll will be randomly assigned a slot.

The poll is in the following link:

https://doodle.com/poll/tyn6cq7at7k3i55i

Good luck!

Hi all,

Since we found some mistakes in the problem spec, homework 4 deadline has been extended to Nov 27. 11:59 PM. The following are a few clarifications of Homework 4:

Exercise 10.1: The second paragraph should be replaced by:

suggest a procedure that relies on $A$ and learns $\mathcal{H}$, in the usual agnostic PAC learning model and has a sample complexity of 

$m_{\mathcal{H}}(ϵ,\delta )\le k{m}_{\mathcal{H}}(ϵ/2)+\lceil \frac{2\log (4k/\delta )}{(ϵ/2{)}^2}\rceil$

where

$k=\lceil \log (\delta /2)/\log ({\delta }_0)\rceil$

Hint: Divide the data into $k+1$ chunks, where each of the first $k$ chunks is of size $m_{\mathcal{H}}(ϵ/2)$ examples. Train the first $k$ chunks using $A$. Argue that the probability that for all of these chunks we have $L_D(A(S))>\underset{h\in \mathcal{H}}{min}{L}_D(h)+ϵ/2$ is at most ${\delta }_0^k\le \delta /2$. Finally, use the last chunk to choose from the k hypothesis that $A$ generated from the $k$ chunks (by relying on Corollary 4.6)

Exercise 12.2: The hypothesis class is defined as $\mathcal{H}=\{\mathbf{\text{w}}\in {\mathbb{R}}^d:||\mathbf{\text{w}}||\le B\}$ and the input domain is defined as $\mathcal{X}=\{\mathbf{\text{x}}\in {\mathbb{R}}^d:||\mathbf{\text{x}}||\le B\}$

Exercise 14.2: Here we assume $ϵ\in (0,1)$

Question 8: Many students realized the problem that if we fix $b=1$ in the decision stump, then the training error of AdaBoost would be very large. If you do not satisfy with the performance, you can implement a decision stump using $b\in \{-1,1\}$, with the following ERM algorithm:

initialize $F^{\ast }=\mathrm{\infty }$

for $j=1,\cdots ,d$

     sort $S$ using the $j$'th coordinate, and denote

          $x_{1,j}\le x_{2,j}\le \cdots ,\le x_{m,j}\le x_{m+1,j}\doteq x_{m,j}+1$

     $F^{+}=\sum _{i:y_i=1}{D}_i$

     $F^{-}=\sum _{i:y_i=-1}{D}_i$

     if $F^{+}<F^{\ast }$

          $F^{\ast }=F^{+},{\theta }^{\ast }=x_{1,j}-1,j^{\ast }=j,b^{\ast }=1$

     if $F^{-}<F^{\ast }$

          $F^{\ast }=F^{-},{\theta }^{\ast }=x_{1,j}-1,j^{\ast }=j,b^{\ast }=-1$

     for $i=1,\cdots ,m$

          $F^{+}=F^{+}-y_i{D}_i$

          $F^{-}=F^{-}+y_i{D}_i$

          if $F^{+}<F^{\ast }$ and $x_{i,j}\ne x_{i+1,j}$

               $F^{\ast }=F^{+},{\theta }^{\ast }=0.5\ast (x_{i,j}+x_{i+1,j}),j^{\ast }=j,b^{\ast }=1$

          if $F^{-}<F^{\ast }$ and $x_{i,j}\ne x_{i+1,j}$

               $F^{\ast }=F^{-},{\theta }^{\ast }=0.5\ast (x_{i,j}+x_{i+1,j}),j^{\ast }=j,b^{\ast }=-1$

I have uploaded corresponding skeleton code of $b\in \{0,1\}$ to the homework code folder. Both implementations will get full credits.

Statistics:

MEDIAN
91.5
MAXIMUM
99.0
MEAN
87.11
STD DEV
9.91

Please find your grade on gradescope, and midterm solution on CCLE. Thanks!

Good job!

Hi everyone,

Registrar Scheduling office just informed me that the final exam (course project presentation) day and time for our class will be Wednesday, December 12 from 3:00-6:00pm. Thanks.

Best,

Quanquan