Update ML discussion section for Day 5

[cgreene]

Add discussion of "when your training data are bad"

http://www.antoniocasella.eu/nume/Xiaolin_2016.pdf https://www.technologyreview.com/s/602955/neural-network-learns-to-identify-criminals-by-their-faces/ https://medium.com/@blaisea/physiognomys-new-clothes-f2d4b59fdd6a

[cgreene]

First 20 mins

[cgreene]

We decided to make this a bit different. The new structure for 2019:

• (prelab) Everyone reads a one-page explanation of both scenarios.
• (prelab) People are assigned by notebook randomization (based on username) into one of two groups.
• (prelab) Two thought provoking items that they can bring to the discussion
• 20 min discussion: One group focuses on the current set of papers (overfitting) / the other focuses on the new set of papers (discrimination).
• 10 mins: Groups come back together to summarize discussion
• Homework should require answering questions from both of sets of exercises.
• Consider streamlining https://github.com/greenelab/GCB535/blob/master/30_ML-III/ML_3_Inclass_Homework.ipynb to make it take less time (20 mins)