TU Wien:Deep Learning for Visual Computing VU (Kampel)/Exam Questions SS21

A collection of different written and oral exam questions. Some questions occured multiple times in different time slots and exam dates.

Written

• What is the difference between a local and a global optimum? Draw a sketch that illustrates the difference. Are local minima a problem in deep learning? Why (not)? What is momentum and why is it beneficial?
• Give a general overview of convolutional neural networks and their purpose, and draw a sketch that illustrates their overall structure (typical layer types and their arrangement). What are the two overall stages of such networks? What is needed to “combine” these stages / to make them compatible and how can this be achieved?
• Assume the following computational graph. Insert digits from your Matrikelnummer into the empty input nodes, going from right to left both in terms of nodes and digits. (Assuming Matrikelnummer 0123456, the values of the rightmost node would be 6, that of the node left of it would be 5, and so on.) Then compute the partial derivative of the topmost node with respect to all input nodes via backpropagation. Write computation node values after the forward pass left of the nodes, local gradients left of the edge connecting the corresponding nodes, and “cached” partial derivatives of the topmost node right to the nodes.

Oral

• U-nets
• Imbalanced Data
• Purpose of loss function
• what measures cross entropy
• Which criteria must the ground-truth labels and predicted class-scores fulfill to support the cross-entropy loss, and how is this ensured?
• Purpose of Convolutional Layers
• What are the two key differences to linear layers, and what motivates these differences with respect to image analysis?
• What is the most popular activation function for these layers?
• What is transfer learning ?
• Regularization, what is it, what are dropout layers, where are dropout layers located within a neural network?
• Image segmentation, what is it, how is it done (U-Net), why does the U-Net have skip connections