|Safe Reinforcement learning |
RL agents are typically trained to maximize a reward signal, which must be specified in advance by a human designer. If the reward function isn’t properly designed, the trial-and-error nature of RL in the real world might lead to unacceptable catastrophes. We'll develop safe RL algorithms by transforming the optimization criterion to include some risk measures (Risk-sensitive MDP or Constrained MDP) and deploy our algorithms in the deepmind safety gridworlds and openai safety gym. Write to me if you are interested in deep RL, safe or meta.
|Dr. Rong Guo|
|Exploring Biologically inspired features for sound event detection in difficult Polyphonic acoustic scenes||Adeoluwawale Adewusi|
|Exploring Blind source separation for Sound Event Detection in Moving Acoustic scenes||Adeoluwawale Adewusi|
|Understanding Semantic Segmentation of Sound Events in Binaural Auditory Scenes||Adeoluwawale Adewusi|
|Semi-Supervised and Data Augmentation Schemes for Polyphonic Sound event detection in Reverberant Binaural auditory scenes||Adeoluwawale Adewusi|
|Training Neural Networks with evolutionary methods instead of back propagation (MSc)|
The standard way for training Neural Networks (NN) is by calculating the gradient of the NN function with respect to the weights/parameters and move along its direction. This method is dominant, mainly due to the fact that the gradient is computationally cheap to calculate. However alternative methods exist. One of this method is the evolutionary approach or gradient free optimization (depending on the community). In this method, one randomly samples various points around the current value of the parameters, and picks a weighted mean of this points as the new point. The advantage of this approach is that it can escape local minima, and that can be applied to cases where the error functions are not differentiable. The goal of the student is to compare these methods in various toy problems.
|Dr. Vaios Laschos|
|Comparing different approaches for creating templates out of brain MRIs (MSc) |
In MRI studies, it is common to create various templates out of the different groups, before one makes the analysis. These templates are some form of averages, which can be used either as a baseline, or as the objects to be compared between groups. The last decade, a method for calculating averages of images has been developed using either the Wasserstein or the Hellinger-Kantorovich distance. Therefore this method can also be applied for MRIs to create templates. The purpose of this thesis is the comparison of traditional methods for creating templates with these newer approaches.
|Dr. Vaios Laschos|
|Training Neural networks that capture the Wasserstein and Hellinger-Kantorovich distance (MSc) |
The purpose of this project is the training of Neural networks that capture the Wasserstein Metric for a specific dataset (MRI or MNIST images). The NNs will be then analyzed and possibly applied to other projects.
|Dr. Vaios Laschos|
|Computational modeling of separable resonant circuits controlled by different interneuron types (Computational Neuroscience, Biophysically Detailed Modeling)||Dr. Christoph Metzner|
|Risk-sensitive deep reinforcement learning||Dr. Rong Guo|
|Deep Learning for Distribution Shift|
|While deep learning is very successful in tasks where training and test conditions are the same, there is still a lot to research on what to do if this does not hold and empirical risk minimization is not optimal. |
Earliest possible start date is April since Thomas Goerttler is on parental leave. In case you are interested you can still contact him before
|EEG: measurements, data analysis, study design |
In our EEG-lab, we offer a variety of EEG (Electroenzephalogram) related topics like (but not limited to)
Realtime synchronization of data:
Using realtime capable equioment (CRIO, NI), incoming measurement data (digital, packeted, UDP) should be "tagged" with absolute timestamps (when the data was collected, not when it arrived), using a reference signal by means of a phase-locked-loop (PLL). After registration the EEG signals should be analyzed (basic methods but implemented in strict RT on an FPGA). Folowing this analysis, the next null-crossing of a likewise extracted stable oscillation should be estimated, send to EEG system and recorded along with the live EEG to check for consistancy. At least good knowledge of "LabView" required.
When measuring EEG, neuronal mass signals are collected at sensors (electrodes) which are subsequently used to deduct brain function. However, in many cases, such analysis of signal levels is highly misleading due to the volume conductance of the human head. The electric signals travel through the partly (skull, white matter) highly anisotropic human head, such that a high amplitude at some sensor does not nescessarily has its origin directly below it. The projection from source to sensors can be described by the so called leadfield matrix using estimates. To reconstruct the signal sources, this matrix has to be inverted which leads to a highly ill-posed problem. Some basic knowledge about electric signals and good math skills would be helpfull for a thesis in this field. EEG studies: possible upon consultation as well as other EEG-related topics
|Dr. Michael Scholz|