Prüfer: Prof. Nöth
Performance Evaluation
„Explain the ROC curve“. (TP-rate, FP-rate, drawing, meaning, Area Under Curve)
„What do we need to be able to use a ROC curve?“ (we need a one-class-yes-or-no-problem, and not a two-or-more-class Problem). That one was very vague and it took me some time until I said what he wanted to hear.
Bayes Classifier
„Explain the Bayes Classifier“. (prior, posterior, Bayes formula, argmax p(y|x))
„How do we get p(x|y) and p(y)?“ (By doing assumptions about the kind of distribution. Then we can estimate the parameters from the training data.)
[I took the gaussian distribution as an example for such an assumption, which led over to…]
Gaussian classifier
How to estimate the parameters
How does the decision boundary look like? (quadratic or sometimes (when?) linear)
„What can we do in order to get rid of the exp(-1/2 * x^T \Sigma x …) part?“ (What are logistic functions, how to formulate a twoclass problem in terms of logistic functions, role of the F(x), decision boundary is F(x)=0)
I also explained how the F(x) will look like in case of Gaussian, and how this explains a linear decision boundary in case of equal variances. Not sure whether he wanted to hear that.
„How large is the covariance matrix of a 100-dimensional-vectorial-data?“ (about ~10000/2 entries (symmetry!), O(n^2))
„Naive Bayes…“ (…assumes independency of the entries, cov-matrix is diagonal, 100 entries)
„Something in between?“ (cov-matrix with only diagonal and some minor diagonals)
„When is this appropriate; why should only some, but not all components be related to each other?“ (time-sampled data or similar)
Unrelated
NN- and kNN-Classifier
„what does the NN do?“
„what requirements for the data?“ (must be normalized, all entries should span the same range)
„how does kNN work?“
„explain the code“ (detailed explanation of the weird matlab syntax needed!)
EM-Algorithm
Summary
Noeth often expected me to continue his own sentences. Questions were extremely vague.
I would definitely not call the atmosphere „kind“. Noeth was pretty picky about minor mistakes, the tone was rather condescending. Nevertheless, the grading was pretty student-friendly and forgiving.