• Tag SS16 course
• Master should represent SS17 from then on
• Repo needs a more appropriate name
• Readme should make all this clear
• New tutors should be invited

Sheet 08 is quite complex. Not only does it require extensive explanations, but additionally the sample solution is quite long and for Python beginners difficult to reason about because of advanced object orientation and class inheritance.

To simplify this I propose an approach more along the lines of my naive MLP implementation. Specifically the first 4 cells:

• One can be (still) be guided from bottom to top, going from a single perceptron over a layer of perceptrons to a multilayer perceptron.
• Classes totally make sense for this. But inheritance is hard to reason about for beginners -- drop it.
• Drop having multiple different perceptrons. To allow this
  • the input is provided as-is, as a list of values, one for each input neuron.
  • Activation functions are not hard coded into the perceptrons implementation but passed to the constructor. This is also more similar to the behavior of common ANN frameworks.

The code is not ready to be deployed. If the general idea is approved I would open a PR changing a couple of things:

• Replace __call__ with activate or something simillar. While pretty, its confusing. For those coming from Java we should stick to something around "classes are called to create instances, instances provide methods".
• Maybe get rid of __traverse__ because its very abstract. Redundant code here is not bad if its easier to understand.
• Consider getting rid of zip and similar shorthands. Some of those, again, are quite hard to reason about and duplicated/longer code is probably preferable.
• Adapt old exercises to new code.

Ping @krumnack @shoeffner @AndreaSuckro