A few students have asked about how late penalties are calculated. Note the assignment instructions:
Late Submission Policy: 20% marks will be deducted from the total for each day late, up to a total of four days. If five or more days late, a zero mark will be given.
This is not a reduction in the cap on possible marks, as some students have asked. It is a daily late penalty deducted from the raw mark as stated.
For assignments 1 and 2 deductions were calculated on a flat percentage basis, meaning that 20 percent was applied to the total marks available for that assignment, multiplied by the number of days late.
For assignment 3, late penalties were calculated as a percentage of the student's total mark. Therefore, for consistency, we have decided to also apply this system to assignments 1 and 2. Some students will therefore see a small increase in their marks for those assignments, which will be entered in the system before final grades are issued.
I hope that clears things up for everyone.
Astrid
Annotated reports have been uploaded to the system under the same link as previously provided for assignment 1:
https://cgi.cse.unsw.edu.au/~cs9418/18s2/view/
Reports should contain marks for each question and some summary comments. For those students who previously asked for a breakdown of marks, please look at your returned assignments first before emailing if there are any issues.
Cheers, Astrid
Marks for the final assignment are now available. Students generally did well on this assessment.
Students did well on the report writing. Where students lost marks on the report this was due to sections being thin or missing important information. Model performance was scaled relative to the best performing model which had an accuracy rate of 94.9 and MNLP of 0.18 on the test set.
Where students received low overall assignment marks, this was almost always due to something going wrong with the model results. Either the predictions were not in the correct format (e.g. no blank lines between blocks, resulting in zero for model performance), or log probabilities were positive (resulting in zero for MNLP performance).
We are endeavouring to get annotated assignments (2 and 3) uploaded for you to see as soon as possible, however exams-related work is currently taking priority for technical staff. I will email everyone once this is done.
Best of luck with your final exam.
Cheers,
Astrid
[Edit: The uploaded reports will contain your model performance as measured on the test set, and comments so you can see the breakdown.]
Hi all,
Many students have asked about their assignment 2 grades. Please note all assignments have gone through a thorough marking process and will not be marked a second time.
In order to provide a better understanding of the grades given, I hope to provide your annotated theory section online shortly so you can see your marks for each section of the theory component (and infer your practical grade), and I will try and give some explanation in response to specific queries where it is possible to do so.
A couple of further comments on the assignment: the theory component asks students to derive certain results - marks therefore cannot be given for providing (or citing) these results without showing derivations.
Good luck in the exam, everyone.
Results for assignment 2 are now available. Below are a few remarks on how students went on the different areas. Cheers, Astrid
Theory
Many students lost marks by not showing all their working, providing final expressions without deriving them.
Notebook
Students generally did well on implementing training code. Some marks were deducted where there was no cross-validation or other parameter search. Marks were also deducted where students used scikit-learn implementations rather than their own code or where they did not implement the model developed in the theory component of the assignment. A few students lost marks due to bugs.
Many students lost marks for prediction due to incomplete code or bugs resulting in no valid predictions being generated.
Submissions that did not generate predictions, or did not produce them in a reasonable time, received zero marks for model performance. A common error was not normalizing (new) test data, however model accuracy was assessed after correcting for this. Students were awarded full marks for model performance where accuracy was above 80 percent. Partial deductions were then made relative to this benchmark, with students achieving 50 percent accuracy receiving half marks for model performance.
For students wanting to see the accuracy of their model on the evaluation test data, you can obtain the test data as per the notebook example for assignment 2 already provided.
Hi everyone, I'm going to move to responding to forum posts at regular intervals. While you are working on assignment 3, I will respond to questions posted on the forum twice a day (around 8am and again around COB). Please post your questions with this schedule in mind. Cheers, Astrid.
Assignment 2 solutions for theory have been posted. An example notebook of how one might implement the model (kindly prepared by Louis) has also been posted.
Info about the exam available under Lectures->Week 13.
As mentioned in the course outline from the beginning of the course and in W12 lecture, there won't be a lecture in W13. Tutorials as usual.
Hi everyone,
I have made available all the questions and solutions for week 13 tutorial and notebook (relating to lecture material from week 12).
That is now all the tutorial materials for the course!
Cheers, Astrid
Hi Students,
My apologies - the tutorial will start a little late today (around 10.15) due to unforeseeable delays.
Cheers,
Astrid
Dear All,
Our last lecture will involve some paper reading (resources on the course website). You are encouraged to look at the material beforehand. However, we will set some time during the lecture for reading and discussions, along with some guidance and explanatory slides.
A small correction has been made to assignment 3: we are now normalising the performance metrics by the length of the sequence. Make sure you have the latest version. Last Update: Friday 12th October, 2018 at 15:41.
Assignment 3 is now available and due on on Sunday 28th October at 23:59 .
Hi everyone,
Be aware that assignment three will be released this week, no later than Friday 12 Oct.
Cheers, Astrid
Hi everyone,
We have a link available where you can login and see your Assignment 1 submission and the marked copy (which has marks for each question and some comments). Link to view your Assignment 1 is below (you can copy and paste this in your browser):
https://cgi.cse.unsw.edu.au/~cs9418/18s2/view/
This is a new tool, so if you have any feedback we'd be happy to hear it.
Also please note that this is probably a "once-off" i.e. we won't return full annotated copies of later assignments, but instead a summary containing the marks breakdown with some comments.
Good luck with Assignment 2!
Astrid
Dear All,
The theory part of Assignment 2 has been updated (by correcting the definition of the covariance (C_{nk}) in question 1(c). Please make sure you use the latest version of the assignment (Last Update: Monday 10th September, 2018 at 09:09).
Assignment 2 is now available and due on Friday September 28th, 2018 at 23:59:59 .
Hi all,
Note that Assignment 2 is expected to be available around Friday 7 September.
Regards,
Astrid
Marks for Assignment 1 are now available. Annotated submissions giving a breakdown of marks for each question will be made available to students in the coming weeks. Late assignments were penalised as per the assignment description.
Solutions for Assignment 1 are now available on WebCSM3. Please note we will not respond to individual queries about grades until students have consulted these first (then students can contact Astrid if still necessary).
Assignment 1 now available and due on Monday 20 August 2018 at 11:59 PM.
Welcome to COMP9418, Advanced Topics in Statistical Machine Learning. This course aims to provide an in-depth study of statistical machine learning approaches, with a focus on methods for learning and inference in structured probabilistic models.
We'll be using WebCMS3 as the platform for content management in our course and a skeleton Course Website is already in place. The Course Outline has more detail on what's in the course and how it will run.
I look forward to seeing you on Monday 23rd July at 12:00 at CLB 5 .