Examiner: Prof. Maier
Give an overview of the topics of the lecture (I drew a mindmap of all lecture topics)
What problems can't be solved by a perceptron? → XOR.
What can we do to solve it?
Universal approximation theorem. How good can we approximate? How can we better approximate?
Why do we need deeper models when we can already model any function with one layer?
Explain backpropagation
What loss function did we use? Cross entropy loss (classification, multinoulli distribution), L2 loss (regression, gaussian distribution)
Short derivation wanted for the loss functions.
We talked about hinge loss in the lecture. Can you say something about this? → SVM
Can you explain some different optimizers? (I talked about Momentum, NAG and Adam)
Name some milestone architectures and say what new ingredient they introduced (I talked about /mentioned LeNet, AlexNet, Network in Network, VGG, ResNet)
Then he wanted me to explain ResNet in detail and asked some more questions about it
Explain LSTM. He wanted me to draw the block diagram
Explain YOLO.
How is segmentation working with deep learning?
Explain autoencoders
Prof. Maier was very friendly. For preparation I wrote my own summary, this was quite helpful