Advanced topics

Question 1

How can BNNs be approximated and thus the uncertainty of a network be modeled?

Answer 1

It can be approximated using methods like Variational Inference, introducing probabilistic distributions over the networks weights, capturing the uncertain of the network

Question 2

Briefly explain main idea and name 2 possible types of semantic representations used in zero-shot learning

Answer 2

The main idea os zero-shot learning is to recognize objects or concepts never seen during the training.
two semantic representations are Attibutes and word embeddings

Question 3

Name four weak annotation-types with which semantic segmentation models can be trained

Answer 3

Bounding-boxes
Scribbles
Single points
Image-level labels.

Question 4

Describe the scenario of domain adaption. Mention which training data is available and what it is the goal.

Answer 4

training data from a source domain is available, but the model needs to generalize well to a different target domain.
The goal is to improve the model’s performance on the target domain using the knowledge gained from the source domain.