MaAnomaly

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

THIS WEBPAGE IS DEPRECATED - please visit our new website

Rohde & Schwarz: Deep feature representation with auxiliary embedding

Thesis by: Christoph Baur
Advisor: Prof. Nassir Navab
Supervision by: Shadi Albarqouni, Athanasios Karamalis

Abstract

While convolutional neural networks (CNNs) show outstanding results in various computer vision tasks, the deep feature representations within these models often lack transparency and discriminative power. By introducing additional regularization to the networks, one hopes to find more robust and discriminative embeddings of the feature representations such that for instance different classes cluster nicely. In fact, previous work successfully employed regularization techniques under supervised and semi-supervised settings and reported improved robustness and better generalization on well known computer vision datasets. The work in this master thesis investigates how different regularization techniques for auxiliary manifold embedding affect the performance of the models - as opposed to previous work - for challenging medical data. Further, the work compares the results using both supervised and semi-supervised training.

Location

Rohde & Schwarz GmbH? & Co. KG.
Berg am Laim
Mühldorfstraße 15
81671 München, Deutschland

Resultant Paper

2017
C. Baur, S. Albarqouni, N. Navab
Semi-Supervised Learning for Fully Convolutional Networks
Accepted to Proceedings of the 20th International Conference on Medical Image Computing and Computer Assisted Interventions (MICCAI), Quebec, Canada, September 2017
A pre-print version is available online at arXiv.
The first two authors contribute equally to this paper.
(bib)

ProjectForm
Title: Rohde & Schwarz: Deep feature representation with auxiliary embedding
Abstract: While convolutional neural networks (CNNs) show outstanding results in various computer vision tasks, the deep feature representations within these models often lack transparency and discriminative power. By introducing additional regularization to the networks, one hopes to find more robust and discriminative embeddings of the feature representations such that for instance different classes cluster nicely. In fact, previous work successfully employed regularization techniques under supervised and semi-supervised settings and reported improved robustness and better generalization on well known computer vision datasets. The work in this master thesis investigates how different regularization techniques for auxiliary manifold embedding affect the performance of the models - as opposed to previous work - for challenging medical data. Further, the work compares the results using both supervised and semi-supervised training.
Student: Christoph Baur
Director: Prof. Nassir Navab
Supervisor: Shadi Albarqouni, Athanasios Karamalis
Type: Master Thesis
Area: Machine Learning
Status: finished
Start:  
Finish:  
Thesis (optional):  
Picture:  


Edit | Attach | Refresh | Diffs | More | Revision r1.5 - 13 Dec 2017 - 19:39 - ShadiAlbarqouni