Cognitive Algorithms Gedächtnisprotokoll exam SoSe18 2 appointment

Multiplichoiche
1. given: picture of simple perceptron
	- classification linear, non linear, depends on f?
	- 0 or 1 hidden layer, unsupervised or supervised learning?

2. your model parameter produces good results but a friend tells you that the model might overfit the data what can you do? write down an idea or a concept concept.
3. see exam SoSe2013 on freitagsrunde task 5.1, same task


True or false and why?
1. more traindata reduces the error
2. linear regression can only solve linear problems
3. ridge regression always out performs linear regression


Some other questions
1. name 2 algorithms you saw in the lectures which can be decompost in eigenvalues
2. given: picture of datapoints in 2 different colors (the data was highly coreelated)
	- draw NCC and LDA seperation
	- which one fits better and why?
3. how does k-means work?


PCA
1. write down PCA in pseudocode and don't forget to comment in every step what will be done


Perceptron
1. explain: "the perceptron solution is based on the initialisation.", how about LDA?
2. given: all formulas of the algorithm and the algorithm it self, some datapoints x1-x4 y1-y4 and w
	- compute w^Tx for x1-x4 which one is missclassified?
	- compute the new w based on the missclassified datapoint
	- show that now all datapoints are classified correctly


NCC
1. given: ||x-w0||² > ||x-w1||² <=> sqrt((x-w0)^T(x-w0))² > sqrt((x-w1)^T(x-w1))²
	- solve it to get 0 < w^Tx - β
2. fill the blank space
	- w =
	- β =


Kernel
1. when to use kernels?


Multilayer Perceptron
1. what do we gain from more hidden layers?
2. explain backpropagation, forward and backward pass


Ridge Regression
1. see exam SoSe2013 on freitagsrunde task 6.3, same task (without equation 9) + which model overfits/undefits the data?