Kolloquiumsvortrag: Dagmar Kainmüller, Max-Planck-Institut für molekulare Zellbiologie und Genetik / 22.01.16
22.01.2016 von 14:15 bis 15:45
Institut für Informatik, Ludewig-Meyn-Str. 2, Übungsraum 2
Titel: Learning Deep Classifiers that integrate Prior Shape Knowledge for Bioimage Analysis
Abstract:
Deep Learning has rapidly transformed the field of computer vision in the past years. To date, deep convolutional neural networks (CNNs) outperform traditional machine learning approaches in a vast number of applications. However, little has been added to the theoretical understanding of CNNs since they have been proposed more than 25 years ago. Furthermore, huge sets of annotated data are necessary for successful training of deep CNNs, while image analysis tasks often come with small amounts of annotated data to learn from — especially in the context of applications in biology.
In this talk I will explore the theoretical relationship between CNNs and decision forests. Decision forests are a hugely popular machine learning tool that can be trained effectively in the face of small amounts of annotated data. They are widely used for image analysis tasks in biology. I will show that decision forests can be mapped to CNNs, which allows for developing novel, easily interpretable deep CNN architectures that can be trained from small amounts of data.