AI-based Applications
AI for Software
Engineering Image and Data Analysis
in Medicine Data Science
for Biomedical Applications AI for
Security
Machine Learning for IT-Security (Prof. Bogdan)
- Objective 1: Efficient accelerators
- Break down Deep Learning to reduce power and memory footprint
- Design efficient circuits for embedded systems
- Objective 2: Secure computation
- Obfuscate computations for privacy in untrusted environments
- Improve resilience of algorithms against intended manipulations
Biomedical Image Data Science (Prof. Sbalzarini)
- Content-adaptive convolutional neural networks
- Develop CNNs for irregular spatial sampling
- Apply to adaptive image representations in microscopy
Image: Representing an image by a content‐adaptive particle representation instead of pixels improves processing
- Hybrid learning of mathematical models from image data
- Develop hybrid learning algorithms using structured sparsity norms
- Apply to learning differential equation models from image data
- Robust learning by design centering
- Replace optimization by design centering when training supervised learners
- Evaluate and compare the approaches
Coding AI: An AI for Code Suggestions (Prof. Siegmund)
- Motivation: Developers frequently trigger code completion functionality of IDEs to get code suggestions for solving
an implementation task- Contemporary code suggestions are inaccurate, very limited in scope, and fail to recognize the programmer’s problem
- Goal: Vastly speeding up development time by supporting the programmer with highly accurate code suggestions for her current implementation task
- Key Idea: Use a generative adversarial network architecture to train a deep generative network such that the generated code suggestions outperform actual suggestions
BigData and Machine Learning applications for advanced image and data analysis in medicine (Prof. Neumuth)
- Within the clinic specific, high-performance AI-based algorithms for data analysis and image reconstruction are required.
- For medical AI, we need: suitable training data, international benchmarking and technical-methodical framework conditions
- Goal: a modular analysis pipeline for multimodal medical data (imaging data as well as other data)
- ML-based preprocessing
- ML-based feature extraction
- ML-based classification
- We will focus on Hyperspectral imaging (HSI), which combines imaging with spectroscopy
Image: The tissue is illuminated, and the reflected light is acquired by a camera and analyzed by a spectrometer. The generated 3D data are called hypercubes with two spatial and one spectral dimension. The technology is used in particular for the classification of tissue (also in situ).
