Registration is closed.
| Time | January 16th | January 17th |
|---|---|---|
| 9:00 |
Technical Session Raik Otto |
Image analysis (Theory) Dagmar Kainmüller |
| 10:30 | Coffee | |
| 10:45 | Technical Session (Hands-on) | Image analysis (Hands-on) |
| 13:00 | Lunch | |
| 14:00 |
Explaining Deep nets (Theory) Gregoire Montavon |
Auto-encoder (Theory) Jonathan Ronen |
| 15:30 | Coffee | |
| 15:45 | Explaining Deep nets (Hands-on) | Auto-encoder (Hands-on) |
The theoretical session starts with an introduction to deep neural networks. A focus is placed on convolutional neural networks for image recognition and techniques to improve their performance on datasets with few labels. Then, we address the question of validation, in particular, how to verify what a deep neural network has learned. We present several methods for explaining deep neural network predictions, including layer-wise relevance propagation (LRP). In the practical session, we consider an image recognition task, train a convolutional neural network on this task, and apply techniques to improve the model accuracy. As a second step, we implement the LRP explanation technique on the state-of-the-art VGG-16 network in order to produce explanations.
About Dagmar Kainmüller
Prior to joining the BIH as the group leader of the 'Biomedical Image Analysis' lab, Dagmar Kainmueller worked as an Independent Fellow at the HHMI Janelia Research Campus in Virginia, USA. From 2013 to 2016 she was an ELBE PostDoc in the lab of Gene Myers at the Max Planck Institute of Molecular Cell Biology and Genetics in Dresden, Germany. Dagmar pursued her PhD in medical image segmentation at the Zuse Institute Berlin and the University of Lübeck, for which she received the BVM Award in 2013.
Image analysis is the process of extracting information from images. In this workshop we will give an introduction into its application to biomedical data. You will learn fundamental machine learning approaches suitable for such data and get to know the basics of how to train neural networks and evaluate their performance. In a practical hands-on session you have the chance to experiment with the U-Net, a popular network architecture in this domain, and to apply it to foreground and instance segmentation. We provide multiple data sets on which you can get a feeling for different hyperparameter such as network depth and size and batch size, as well as the various challenges inherent to these methods.
Currently, I am working on a startup trying to bring machine learning models in cancer genomics to the clinic. I founded www.arcas.io together with my supervisor Altuna, and two postdocs from my lab, Vedran and Bora, and we're funded by the BIH Digital Health Accelerator. I got my M.Sc. in electrical and control engineering in Trondheim in 2010, worked as a programmer and data scientist around Europe until 2013. Then I moved to NYC and worked as a computational social science researcher at NYU. I moved to Berlin in 2015 and submitted my Ph.D. thesis in computational biology in September 2019. I've taught courses and seminars on Linear Systems, Machine learning on social media data, RNA-sequencing analysis, and Single Cell RNA-seq analysis.
In my presentation, I will introduce autoencoders and some of the main autoencoder variants that are being applied today. I will show how these can be used to integrate multi-omics data in cancer research. Finally, there will be a hands-on session where the participants will implement autoencoders in TensorFlow 2.0 and Keras, and learn about some practical considerations when using them.
Seminar room 104
IRI for Life Sciences
Philippstraße 13
https://goo.gl/maps/9LUWXKXj6pv
Basic programming skills in python 2.x or python 3.x are highly recommended. In case you are not familiar with Python, please do consider the acquisition of basic Python skills for example by working through the first two chapters of this tutorial
https://anandology.com/python-practice-book/index.html
or any other tutorial book or your choice.
You can bring your own laptop to learn the setup specific to your operating system. If you do not have a laptop, please send a short note to Manuela Benary via compcancer@charite.de.
Please make sure that you have access to the Eduroam WLAN network.
For those of you that utilize the Windows operating system, please acquire the 'PuTTY' software, for instance from https://www.ssh.com/ssh/putty and work through the basic tutorials of how to connect to a remote server. We will walk you through the PuTTY process during the workshop, but in order to focus on Deep Learning aspects, preparation is recommended.
