Post mortem, autumn 2016

Dear diary, A few days ago (6, tbh) we finished the autumn course. 'twas mostly terrible. Here are a few 'hits' and 'misses' i made.

Most people (70%+, self-assessed) are able to solve the standard "classify image/text"/"generate sequence"/whatever problems
Some people (25%, self-assessed) are merely familiar with some keywords, which may or may not help them later in their research/job. I can't claim which extent of that was due to poor teaching and which was them not investing the time/effort.
There are roughly 4-5 students who dove into the topic and went on studying it beyond the basic curriculum. That's awesome. But these are the people who would have done well even if i just told 'em "Guys, you may want to hack some deep learning. Here are some useful links" (no data to back that).
projects failed (as an incentive). The only people who did them would have done them anyway.
we really messed up with the grading system. The result was overcomplicated since we didn't discard the unhelpful rules

maybe split week1 into two weeks: one about the core principles (& manual backprop) and one about the tricks (& theano)
[several student opinions] maybe need a dedicated [maybe optional] material on manual backprop
[several student opinions] add some words (and mb optional assignment) about other approaches to NN training. Contrastive divergence & RBM, DNI, irprop, alternative backprop techniques.
[no data] generative models & lidl's seminar should be given after the basic bayesian methods seminar cuz of some dependencies
[one student opinion] maybe shift deep RL part to an earlier timeslot and push generative models further?
[jh self-assessment] probably need to cover autoencoders a bit earlier. Perhaps near the advanced vision section.
[jh self-assessment] maybe add the differentiable data structures stuff from dl-course
maybe rearrange lectures to mix the math-heavy stuff with a case-study stuff

Surprisingly, student's own opinion supports informal lecture style. I still wonder if they are as efficient as normal ones, but at least they're not repelling ppl. There are some 2 Strong No's though (vs 11 strong yes).
compile PDFs from the offset
add some recommended reading after the lecture for those top10% self-motivated ppl
[e. sokolov advice] add even more explanatory materials on convolutions @week2
[jh self-assessment] lecture 4 is okay, but probably need more interactive part to facilitate RNN understanding
[jh self-assessment] lectures 6, lecture

Students support the "What's wrong with this network" idea. Need to make more of these.
May want to showcase some of the best homeworks [no data to back].

No data to back it, but i believe that "Fill in your code" style homeworks ought to be mixed with "Write me from scratch" ones to make students understand the whole pipeline.
HW0 should remain simple to incentivize doing bonus stuff. Perhaps incentivize bonuses even more explicitly.
splitting old mnist-hw into two was [seemingly] a good idea, both appeared as tasks of moderate complexity [from student responses]
[grading table] base assignments (hw0-hw4) were accomplished by the majority of students, while the second part went largely unnoticed
fix the hw5 mess up :)
Publishing almost-solved HWs should be a no
The same is valid about HW lore parts

maybe the current grading system actually discourages people from doing the second half of homeworks
should think 5 freaking times before introducing a new grading rule mid-course
moar autotests & checkers
could've been much worse :)

Provide feedback