Dear all,
you are cordially invited to join the master thesis defense of Jens von der Heide taking place on the 22nd of February at 13:00 (s.t.). You can find the Zoom credentials below. The topic is as follows:
Exploring the Impact of Varying Class Distributions on Concept Embeddings Generated with Net2Vec
Since the victory of Deep Blue against the World Chess Champion Garry Kasparov
in 1997, the use of deep learning models has reached many domains of
human life. Yet, especially in sensitive areas like medical diagnoses, more comprehensibility
of the results is needed to increase the patient acceptance. The
discovery of more comprehensive deep learning classifiers is currently one of the
most active endeavors in the field of machine learning, and specifically in deep
learning. The framework Net2Vec seeks higher explainability by decoding semantic
concepts in neural network layers. The original authors used the entire
BRODEN dataset, a large image dataset densely labeled with concept segmentations
and classifications, in their experiments, without considering relations
between target classes. In order to extend this research, this thesis examines the
influence of different training data on the original Net2Vec results experimentally.
Particular semantic concept combinations of BRODEN are used as target
classes in various Net2Vec runs. The experiment results show the existence of
semantically related sub-concepts encoded in the convolutional layers. Greater
similarity of the target classes employed in the training lead to more expressive
and more unambiguous sub-concepts. These findings enable deep learning
model decisions with respect to semantic concepts to be more comprehensive.