10-606, 10-607: Math Background for Machine Learning

Hi all,

Thanks for submitting all of your reports. They look great!  [If anyone has not yet submitted, you should avoid looking at this until you do!].

I've run the final numbers, and plotted the results below. Interestingly, it seems as though overfitting to the dev set ended up being somewhat of a problem. About 20 people got dev MSE < 1.0, and most of these also had test MSE < 1.0. For those with dev MSE > 1.0, on the other hand, most people had a much higher test MSE. Even among those people with dev MSE < 1.0, there was a general trend that the smaller the difference between train MSE and dev MSE, the better the final result. (This can be seen more clearly in the second plot, which is on a log scale).

In lieu of solutions, I've posted the report by Dan Mahr to the Piazza resources, which does a really nice job of laying out the stages of data exploration, designing transformations, and gradually improving the model. (He also achieved the lowest test MSE!). I've also uploaded the true transform() function that was used in generating the data. 

If you want to check your own result, I think you should be able to log on to repl.it and re-run the unit tests (after first clicking run). Just make sure NOT to unsubmit / resubmit, or it will appear as late. [There are a couple that were giving me problems, but don't worry if you run into this; I just haven't fixed it yet].

Good luck on the final exam!

Hi all,

To make sure everyone knows this: the exam logistics will be the same as last time. That is, we'll release the exam at 1:30p ET, and you'll have until 11:59p ET to submit on Gradescope. We don't expect the exam to take nearly that much time, but the extra time allows flexibility for students in when to write up the solutions, and also lets students submit early to allow time to resolve any problems with submission. As always, please don't leave it to the last minute to submit; the site closes at 11:59.00p, and it's your responsibility to submit early enough to resolve any problems you encounter.

Hi all,

We've gone ahead and released HW2. It's a shorter assignment than HW1, since there's not that much time left in the semester; but as always please do not leave it to the last minute. See https://canvas.cmu.edu/courses/604/assignments for details.

Hi all,

A reminder that there is no class on Wednesday 11/22 due to Thanksgiving break. Next class is Monday 11/27; see you then!

Hi all, 

I have uploaded solutions for the minimum relative entropy short assignment. Please find them in the resources section of Piazza. Also, please note that I have not posted the grades to Canvas yet, but will try to do so soon.

Thanks,

Dallas