Link that might be useful for PCA:
http://setosa.io/ev/principal-component-analysis/

Should talk about covariate shift more:
https://www.analyticsvidhya.com/blog/2017/07/covariate-shift-the-hidden-problem-of-real-world-data-science/

Date-time is a good example of feature engineering!

Should add polynomial fitting as simple, example of feature engineering.

Add mark expectations and plagiarism definition to opening of ML1

Students don't know parameteric vs. non-parameteric
Students think random forest is slow! (because their implementations suck?)



2018:
L01: First lecture I feel really worked, though they were a bit quiet. Did end 5 minutes early, but that's fine given the number of questions afterwards.

Thing to add:
* Classification - To which category does this data point belong?
* Regression - Given this input from a dataset, what is the likely value of a particular quantity?
 * Clustering - Which data points are similar to each other?
* Dimensionality reduction - What are the most significant features of this data and how can these be summarised?
* Semi-supervised learning - How can labelled and unlabelled data be combined?
* Reinforcement learning - What actions will most effectively achieve a desired endpoint?


L02: 25 minutes early. Need more! Think what I have is mostly good though; could do with more interaction however. Spurious reference to probability in summary, no further reading, needs examples. Examples plus regression would work, plus discussing discrete inputs. As an aside, was in lab immediately trying to git pull from my webserver and couldn't - took me half an hour to figure out that my ssh configuration had been deleted. I assume BUCS are the guilty party.

L03: Timing good (well 10 minutes early), but contains lots of material that should go into the above. One student objected to speed I was talking (too fast). Discovered panopticon wasn't making videos visible! Think an animated walkthrough would help hammer it home. Another demo would be good - they seemed to perk up for Kinect. Make Wine random forest slide silly, just because. Could be clearer with regards to fitting a distribution then doing info gain. Still no interaction:-( Didn't meantion Kinect going continuous for part labelling, which would add to story.

L04:
Right length, good content - think this one works. Slides a bit ugly/text heavy however so could do with beautification.


L09:
15 minutes early, but think that was ok - 42 slides! Seems to work, first half is actually quite nice, but second half still rough and in need of more diagrams for clarity. Also just some more fun images and better Bayesian demo! Notes in Latex cover everything missing!


L10:
10 minutes early. PCA: Question about need for basis vectors to be orthogonal - should improve that. Think a visualisation of the superconductor demo would help. Examples of feature engineering are whats needed though - maybe three slides, one example each, from different domains? SIFT, word vectors and financial modelling maybe?


L12:
Need to explain the relationship between a factor graph and the Bayesian network backwards - the example where you draw the RV circle around the factor and factor graph RV to explain how the factor graph equations become the BP equations.



2017:
L01: Finished 15 minutes early, or 10 minutes early given questions at end, so can add some more. Students were very quiet - didn't want to interact. Lots of questions about Jupyter - maybe add 10 minutes of slides on it?

L07: Finished 20 minutes early, even walking though code. Extra feature trick to make circle work was not clear to them. Need the polling demo:-/ Probably a bit more content as well. Was lossing voice...

Forgot to discuss stupid baselines!

L08: Just needs more of everything. Some of: https://arxiv.org/pdf/1606.06565.pdf
https://lukeoakdenrayner.wordpress.com/2017/12/18/the-chestxray14-dataset-problems/
random noise vs structured noise - should mention this terminology.


L09: Finished 15 minutes early, disjointed, needs serious work.

L10: Finished 20 minutes early. Pretty bad, not sure I gave them all of the intuition for PCA and had to adhoc examples of feature engineering.

L11: Finished in 30 minutes, far too abstract past the start; crap basically. Didn't talk about collapsing loops in factor graphs. Be clearer about unary/pairwise.

L12: 10 minutes early, which is much more acceptable. Needs an explicit example and better stepping through - complete. Maybe something using colour?

L13: Finished 25 minutes early. Could have added LDA. Live demo would be good.


