Institute of Neural Information Processing

Thesis Topics (Bachelor and Master)

We offer methodologically diverse thesis topics in the fields of machine learning, cognitive modelling, computer vision, computational neuroscience, behavioral neuroscience and neuroeconomics. You can either work on one of our currently open topics or you can also develop your own topic suggestion. Depending on the research field and topic, the thesis can have a more theoretical, practical or experimental focus.

Open-ended Learning
Learning based on the interaction with the environment

Exemplary Topics

One important problem to tackle in reinforcement learning is directed exploration. There are different approaches in the reinforcement learning literature that formalize the concept of curiosity in order to capture the idea that some aspects of the environment should be considered interesting to an agent. Depending on the approach, one might focus on prediction error, learning progress, or familiarity with the environment to quantify the intrinsic reward.

Example: In Pathak et al. (2017), the authors explore the use of prediction error between subsequent states as intrinsic reward. More precisely, the prediction error is calculated in latent space, in order to be able to focus on more abstract states differences. However, this error cannot be used to train the embeddings  themselves, because this would result in collapse (one could map everything into one latent state and get zero error). Instead, they are trained using an inverse model that predicts the action connecting two subsequent states. What happens if we replace the square prediction error with a learned distance function between the latent representations of states, based on the amount of steps between them?

Contact: Zeqiang Zhang

Traditionally, policies in reinforcement learning are trained to solve one particular problem, e.g. reaching a fixed goal state in a mace. More recent approaches make policy networks goal dependent instead of learning separate policies for multiple goals.

Contact: Fabian Wurzberger and Zeqiang Zhang

Example: Combining latent world models in simple environments hierarchically.

Contact: Fabian Wurzberger

Example: Use and fine-tune pre-trained navigation models (e.g. general-navigation-models.github.io) on robotic hardware, train goal-conditioned policies, etc.

Contact: Fabian Wurzberger

Example: Currently most model-based reinforcement learning algorithms train agents episodically. As resetting may seem unnatural for agents in real-world scenarios (robots, animals, humans, etc.), it is important to understand the implications of this choice.

Therefore, in this work, we want to compare world models learned with and without resetting.

Contact: Fabian Wurzberger

Representation Learning
Learning latent representations by supervised and unsupervised learning in neural networks

Exemplary Topics

Example: Comparing reconstruction-based methods, such as Variational Autoencoders (Kingma, Welling (2013)), to joint embedding architectures (e.g. Assran et al. (2024), I-JEPA). 

Contact: Sebastian Gottwald

Example: Many unsupervised contrastive representation learning methods such as SimCLR (Chen, Kornblith, Norouzi, Hinton, 2020: SimCLR) can be viewed as energy-based models (LeCun (2021), Bardes, Ponce, LeCun (2022), Dawid, Lecun (2022)). How do contrastive methods compare to regularized methods? What happens when the energy is learned as well?

Contact: Sebastian Gottwald

Example: Comparing different methods to learn discrete hidden representations, such as VQ-VAE, Dreamer-style encoders, or other methods, and how they compare to their continuous counterparts. 

Contact: Fabian Wurzberger and Sebastian Gottwald

Example: The paper Balestriero, Pesenti, LeCun (2021) indicates that deep neural networks, at least the specific architectures studied there, do not interpolate. What about other types of architectures, such as Transformers or layers that specifically interpolate latent states, such as Verma et al. (2019).

Contact: Sebastian Gottwald

Example:

Contact: Sebastian Gottwald

Contextual computation
Topics on incorporating additional, contextual information into the processing of deep neural networks.

Exemplary Topics

How data is to be interpreted and processed oftentimes should change in light of context. For example, what kind of action is performed with a knife not only depends on the knife itself, but as well on the person(s) and objects involved and the scene as a whole. Overall, evaluation context in a distributed representation requires many-to-many interactions, as, e.g., found (self-)attention mechanims of transformer models. The result is non-localized, many-parameter model components to establish long-range interaction.

In this work different mechanisms of how to incorporate long-range, contextual signals within local processors will be investigated. The investigated mechanisms will be based on promising previous work that draws inspiration from the brain. The choice of dataset and task domain is rather flexible, as long as they provide a potential for contextual information integration.

Literature:

Contact: Daniel Schmid

Medical experts routinely screen images for signs of abnormality indicative of a disease. Experimental evidence shows that these experts are able to make a diagnose above chance level even after screening images for only a few seconds. Such visual inspection utilizes global ensemble scene statistics, or “scene gist”, to provide contextual guidance information. This information alone, however, does not equip an observer to precisely localize the location and feature composition indicative of the evidence. Sch global contextual information might then help to guide attention selection mechanisms to constrain target selection via informed contextual perceptual search mechanisms.

The goal of the thesis work is to compute contextual gist information and incorporate this in machine vision approaches, such as deep neural networks for image processing and medical target detection. In this work, neural mechanisms are integrated intro pre-trained CNNs to compute scene gist for contextual guidance in radiological images. Computational nodes in hierarchical CNN architectures will be advanced by integrating local bottom-up feature extraction mechanisms with top-down modulatory contextual fields. These mechanisms are targeted to learn their combined integration in counter-stream networks by employing two-point information integration units. In all, these investigations advance generic CNN architectures by integrating different data streams that operate at different spatio-temporal scales and feature types. Prospectively, such model network aims at performing similarly like a human medical expert.

Literature:

Contact: Daniel Schmid

For many real-world problems it is necessary to combine different streams of sensory information, such as, camera-based image or video data, acoustic signals, tabular or text-based data. Designing specific network building blocks and deep neural architectures which are capable of such information fusion to increase task performance is therefore an important goal.

Here, data fusion for the case of mutli-modal action segmentation in should be explored. The investigation can be based on several annotated datasets in which humans are performing actions and for which video recordings and eye-tracking data exist. To incorporate the eye-tracking data existing approaches will be extended for this new modality and subsequently evaluated on the respective datasets.

Literature:

Contact: Daniel Schmid

Neuromorphic and bio-inspired computation
Investigations with novel, bio-inspired compute hardware, and models.

Neuromorphic Computation: Exemplary Topics

Neuromorphic hardware provides great promises regarding resource efficiency and real-time capabilities. Given its brain-like decentralized asynchronous processing paradigms, neuromorphic processors provide a natural fit for (deep) neural networks. So far there is no universally agreed upon neuromorphic architecture and hardware specification. Whether and how neural networks can be described on a specific processor is therefore dependent on the respective hardware. Specifically, not all hardware permits for on-chip learning of neural networks. An alternative is provided by conventionally training neural networks and then porting the learned parameters to a matching on-chip network architecture of the model.

Within this topic the goal is to inspect different such porting schemes onto the SpiNNaker neuromorphic architecture and to evaluate the networks' performance on standard benchmark datasets.

Literature:

Contact: Daniel Schmid

Classical rate-based neural networks require highly redundant processing of information in parallel, which is carried out with energy-demanding general-purpose graphics processing units (GPUs). Spiking neural networks offer new avenues to encode information in sparse, precisely timed, and potentially energy-efficient ways. Yet, while a highly-performing traning procedure (stochastic gradient descent) exists for rate-based neuron models, a training procedure that would work similarly for spike-based neurons hasn't been identified so far. Another option to bring the latest successes of deep learning to spike-based neural networks consists in a rate-to-spike conversion. There, pretrained rate-based networks are mapped onto spike-based architectures. Different convesion schemes exist - all with the goal in mind to provide the best possible match of the spike-based network to the performance of the rate-based one.

Within this topic the goal is to further the understanding about rate-to-spike conversion methods regarding their performance and generality. To this end, different such conversion schemes and network architectures need to be inspected and evaluated with respect to the networks' performance on standard benchmark datasets.

Literature:

Contact: Daniel Schmid

Visual motion integration is the process of combining evidence about detected motion from a distributed code to accurately represent moving objects. Much about how such integration can be computed has been learned from the brain and utilized for machine vision applications. Beyond such algorithmic inspiration, also on an implementational, or hardware, level improvements towards more energy-efficient, real-time-capable artificial systems can be made by taking inspiration from the biology. Neuromorphic processors and vision sensors do so by incorporating principles of event-based, decentralized and asynchronous information processing and transmission as found in the brain.

In this work an existing biologically inspired model of visual motion processing will be mapped and implemented on the SpiNNaker neuromorphic hardware and evaluated using real-world motion signals recorded from a dynamic vision sensor (DVS).

Literature:

Contact: Daniel Schmid

Bio-inspired Computation: Exemplary Topics

Different from most of the common deep learning techniques, real-world organisms need to (re-)act (to/)in continuously changing environments. Processing information in and choosing actions in sequence is, thus, a distinctive property and required for autonomous deplyoment of machines into the real world. How such sequential processing can be made configurable and applicable to artificial neural networks is therefore of great interest.

The goal of this investigation is to further our understanding of how different mechansims can be utilized in bio-inspired artificial neural networks to implement properties of sequential processing, and learning.

Literature:

Contact: Daniel Schmid

To express sequential computation neural patterns need to likewise evolve in a sequence. Such dynamic evolution of neural patterns requires also a clear conceptual understanding of how to steer these dynamics and how to interface with the network in order to retrieve useful outputs and provide inputs in a timed fashion. Furthermore, to generalize well beyond individual sequences, modularization of sequential evolution into computational motifs is a desirable property of such networks.

In this investigation a functioning prototype should be created, that serves as a proof-of-concept. The aim is to show how a biologically inspired dynamical controller can steer and interface with a dynamic neural network and how this approach facilitates modularization and re-use of dynamical motifs.

Literature:

Contact: Daniel Schmid

Theoretical Topics
Various more theoretical topics

Exemplary Topics

Example: Confirmation of the results in the paper Shwartz-Ziv, Tishby (2017), Opening the Black Box of Deep Neural Networks via Information using other mutual information estimates, such as MINE. What happens in different architectures, such as Convolutional Networks or Transformers?

Contact: Sebastian Gottwald

In general, standard neural network architectures and loss functions lead to averaging output behavior when confronted with stochastic inputs, resulting in many-to-one mappings, mostly due to two reasons: 1. similar inputs are more or less mapped to the same lower dimensional latent which translates to a single output, 2. the usual loss functions (cross entropy, square error, etc.) are optimal at the corresponding expected values.

Example: How can energy-based models be used to represent uncertainty? How do modern generative models allow for many-to-many and one-to-many mappings?

Contact: Sebastian Gottwald

Example: 

There are several attempts to make use of geometries in deep learning other than the Euclidean one. For example, hyperbolic geometry has an innate hierarchical structure (see, e.g., the Poincaré ball) which is why it might be suitable to augment neural networks with, in order to create/increase the tendency of hierarchical abstractions.

Example: Figure out a way to efficiently use hyperbolic geometry in neural networks and analyze whether this creates an inductive bias towards more hierarchical representations.

Contact: Sebastian Gottwald

Neural networks generally are not aware of their own uncertainties. Even for unseen input data a neural network typically has a single deterministic output. If not generalizing well, this can result in unexpected outcomes (e.g. "hallucinations" in LLMs) which is a problem for the application of neural networks in critical scenarios, such as medical applications, traffic, etc. Moreover, in the design of intelligent agents (e.g. in reinforcement learning), higher uncertainty for novel and less-common inputs would be desirable, e.g. for exploration or continual learning. Therefore, neural network architectures that allow the reporting of their own uncertainties is very valuable.

Getting started: For example, study and implement the approach by Kirchhof et al. (2024) which is based on pretraining. Or consider uncertainty estimates based on the variance term in bias-variance decompositions such as Gruber, Buettner (2023).

Contact: Sebastian Gottwald

Past Theses

Abgeschlossene Bachelor- und Masterarbeiten

  • Arends (MSc.): Psychophysiological classification of emotional and cognitive load states and evaluation of various machine learning algorithms
  • Betgov (MSc.): Multi-label (category-aware) semantic edge segmentation in agricultural applications
  • Braig (MSc.): Generation of synthetic order data for the vehicle touring problem
  • Epp (MSc.): Modellierung eines Reinforcement Learning Ansatzes zur Handlungsabsichtserkennung basierend auf psychologischen Modellen
  • Franke (MSc.): Improving Pedestrian Detection and Evaluation by Utilizing Image Segmentation
  • Freudenreich (MSc.): Automatic localization quality evaluation of a mobile robot
  • Fundel (MSc.): Automatic bat call classification using transformer networks
  • Hegde (MSc.): Data-Free Quantisation and Pruning of Deep Neural Networks
  • Hickmann (MSc.): Reinforcement learning with variational quantum algorithms
  • Horvath (MSc.): Improvemment of a GNSS-Position by using Image Segmentation for Outdoor-Autonomous Systems
  • Janzen (MSc.): Angewandtes Matching von Anfragen mittels kombinierter Analysemethoden für Plattformen
  • Klein (MSc.): Learning from Human Feedback: Implementation and Evaluation of TAMER with Counterfactuals in a Social Robot
  • Kramer (MSc.): Sensor fusion based on recurrent neural networks for indoor navigation
  • Lisogorov (MSc.): Annotation Efficiency in Neurosurgical Instrument Tip Detection
  • Lochner (MSc.): Evaluating GAN Architectures for Time Series Synthesis
  • Meuser (MSc.): Erstellen einer wissenschaftlichen Vorgehensweise zum Bestimmen der Relevanz von Monitoring KPIs
  • Neumann (MSc.): Erkennung und Füllstandsmessung der Zonen in Kühlgeräten mithilfe von Computer vision und künstlichen neuronalen Netzen
  • Rapp (MSc.): Schädigungsanalyse von QFN-Komponenten mittels neuronaler Netze unter Verwendung von In-situ-Sensordaten aus Lebensdauertests
  • Rost (MSc.): Double Descent – The Behavior of Neural Networks Depending on Their Number of Trainable Prameters
  • Sattler (BSc.): Evaluation und Bewertung synthetisch erzeugter Zeitreihen
  • Schuster (MSc.): Erstellung eines Konzepts für die Anbindung und Evaluation eines intelligenten Modells für Machine Learning
  • Ünlü (MSc.): Development of Data-driven Strategies to Improve Customer Approach: application of Data Mining Methods to Reduce Customer Churn using the Case of Financial Service Provider
  • Wehle (MSc.): Tracing einer Lichtstruktur im dynamischen Fahrbetrieb zur 3D-Rekonstruktion
  • Weigerstorfer (MSc.): Crack Detektion mithilfe von Deep Learning für den Lochweitenaufweitungsversuch
  • Zhang (MSc.): Learning skills without rewards in the Non-IID reinforcement learning environment

  • Akram (MSc.): Video Sequence Modeling with Slot Attention
  • Bernard (MSc.): A context based approach for multi-labrl classification of movie genres from plot synopsis using BERT
  • Brieger (MSc.): Learning Task-Relevant Representations for Real-World Reinforcement Learning with Selective Contrast
  • Eberhard (BSc.): Maschinelles Lernen zur Erkennung von "Fake News"
  • Geisel (BSc.): Biosignal Based Emotion Recognition by Machine Leanring Algorithms
  • Graf (BSc.): Tiefe neuronale Netze zur Erkennung von COVID-19 Erkrankungen in medizinischen Röntgenbildern
  • Gupta (MSc.): Perceptual grouping mechanisms in neural networks using gated recurrent units
  • Happacher (MSc.): Elektrische Batteriezellenmodellierung auf Basis neuronaler Netze
  • Hoffmann (MSc.): Multimodal synthetic time series generation with Generative Adversial Networks
  • Hofstäter (MSc.): Compressing information through a communication bottleneck – Taking a closer look at the Shared Workspace Vision Transformer
  • Jung (MSc.): Dynamic Asset Allocation using Reinforcement Learning
  • Kaur (MSc.): Decision making under time varying resources modelled through the information theoretic bounded rationality model
  • Köhler (MSc.): Evaluating representational similarities in deep neural network architectures trained by Q-AGREL
  • Kohler (BSc.): Datenanalyse zum Transferlernen bei der Schmerzerkennung
  • Kulkarni (MSc.): Pedestrian Awareness Detection and Evaluation of its Influence on Trajectory Prediction Performance
  • Lauber (MSc.): Synthetic Training Data for Object Detection in an Industrial Environment
  • Lell (MSc.): Monte Carlo Droput in Ensemble Fusion
  • Ludwig (BSc.): Exploration Strategies for Sparse Reward Environments in Reinforcement Learning
  • Metwaly (MSc.): Robustness of a biophysical model for approximating the backpropagation algorithm
  • Nandagudi (MSc.): From depth images to safe motion planning using contrastive deep reinforcement learning
  • Pavel (MSc.): Echtzeit-Emotionserkennung in virtuell-immersiven Trainings- und Kollaborationsumgebungen mit empathischen sozialen KI-Agenten
  • Qi (MSc.): Evaluating biologically inspired initial layer for convolutional neural networks
  • Pauly (MSc.): How to incentivize efficient learning choices in digital learning environments? A reinforcement learning approach applied to an educational game
  • Raichur (MSc.): Image-/IMU-based Deep Multimodal Information Fusion for Pedestrian Self-Localization
  • Riken (MSc.): Person independent pain detection: a domain transfer analysis based on physiological signals
  • Schneider (BSc.): Radial Basis Function Networks for the Estimation of Pain Intensity
  • Schwegler (MSc.): Learning domain specific word embeddings for product categorization
  • Seid (BSc.): Generierung von synthetischen Daten unter Verwendung generativer adversieller Netze
  • Steinert (MSc.): Personalisation Techniques Based on Physiological Information in a Pain Intensity Recognition Scenario
  • Steur (MSc.): Next-Generation Neural Networks: Capsule Networks with Routing-By-Agreement for Text Classification
  • Stockdreher (MSc.): Artificial Intelligence in Intensive Care: Prediction of Critical Patient Conditions and Causality Detection
  • Wehinger (MSc.): Deep Generative Replay with Regularized Autoencoder forLong Task Sequences
  • Zhang (BSc.): Extraktion von Wartungsinformationen aus der Herstellerdokumentation zur Erstellung von Wartungs- und Inspektionsplänen mt Hilfe von KI-Methoden
  • Zhang (MSc.): A Benchmark Environment for Non-IID Continual Reinforcement Learning
  • Zhou (MSc.): Domain Transfer Learning for Multi-Object 3D Scene Decomposition
  • Zhong (MSc.): Efficient Operations for Lightweight Deep Neural Networks
  • Zuccarello (BSc.): Data Transfer in Zero-Shot Learning

  • Bharadway (MSc.): Spiking Network Model of Hippocampus to Store and Recall Individual Patterns
  • Birk (MSc.): Active Learning for Speech-Based Emotion Recognition
  • Bayirli (MSc.): Analyzing Adversarial Robustness of Neural Networks with Randomization Layers
  • Dalferth (MSc.): Visuelle Analyse von Verschleißarten mit neuronalen Netzen am Beispiel von Zerspannungswerkzeugen
  • Feix (MSc.): Parkinformationssystem München – mit Datenanalysen und Prognosemodellen zu vorausschauenden Mobilitätskonzepten
  • Fischer (MSc:): Methoden der ordinalen Klassifikation zur automatisierten Schmerz-intensitätsschätzung
  • Grüninger (BSc.): Multimodale Schmerzklassifiktion mit OpenSSI
  • Günther (BSc.): Visual Analysis of Printed Circuit Boards
  • Harsch (BSc.): Classification using confidence estimation based on Monte Carlo dorpout in multiple classifier systems
  • Hickmann (BSc.): Iterative Autoencoder Networks for Signal Denoising
  • Holve (MSc.): Personalized emotion recognition based on EEG
  • Klimmek (MSc.): Faster Training of Deep Neural Networks for Noise Reduction: A Meta-Learning Approach
  • Knoblauch (BSc.): Reinforcement Learning – Taktische Entscheidungsfindung im Bereich autonomes Fahren
  • Kögel (BSc.): Myogramm-basierte Gestenerkennung mi6els (efer neuronaler Netze
  • Kreuzer (MSc.): Deep convolutional and LSTM networks on multi-channel time series data for gait phase recognition and stride length prediction
  • Lenders (MSc.): Forecasting electrical load: an ensemble approach
  • Link (MSc.): Conception and development of a machine learning framework for laboratory automation solutions
  • Lutz (BSc.): Assoziative Verarbeitung natürlichsprachlicher Inputs am Beispiel eines Taschenrechners
  • Narang (MSc.): Entrainment of brain oscillations via amplitude-modulated transcranial alternating current stimulation (AM-tACS) modulates visual perception
  • Rausenberger (BSc.): Methoden zur Clusteranalyse von Sequenzen für Tastendruckdynamiken
  • Rottach (MSc.): Road boundary detection based on clustered radar data and with neural networks and receptive fields
  • Sawczak (BSc.): Ensemble Learning with Echo-State Networks for Regression Tasks
  • Schmid (MSc.): Phishing Detection with Modern NLP Approaches
  • Schmid (MSc.): Recurrent neural networks for learning incremental grouping tasks
  • Schwarz (MSc.): Improving planning in model-based reinforcement learning with adaptive action priors
  • Schweiger (MSc.): Deep Learning Verfahren zur Auswertung vorverarbeiteter Bilddaten für Aufgaben der ultraschallbasierten Objektklassifikation in Fahrerassistenz-systemen
  • Sedelmaier (MSc.): Deep learning based tone mapping for infrared images and videos
  • Usmani (MSc.): Training convolutional networks with event based visual input
  • Vendt (MSc.): Image to Image Transformations using Conditional generative Adversarial Networks
  • Weber (MSc.): Optimising Convolutional Neural Networks for Event-Based Data
  • Widiyam (MSc.): Generating original artwork using a generative adversarial approach for zero-shot learning
  • Wildberg (MSc.): Residual Echo Suppression based on Neural Networks
  • Yu (MSc.): Uncertainty calibration for a preprocessing enhancement and downstream detection pipeline
  • Zanotti (BSc.): Neuronale Netze als dynamische Ersatzmodelle für die MKS Simulation
  • Wally (BSc.): Machine learning and pattern recognition with applications in affective computing
  • Abosamaha (BSc.): Learning to play Tetris using reinforcement learning
  • Ismail (BSc.): Deep convolutional neural network for age estimation based on pertained models
  • Mahmoud (BSc.): Applying Deep Q-Networks (DQN) to the game of Tetris using high-level state spaces and different reward functions
  • Samaha (BSc.): Binary classifier models: Training on phasic electro data and testing on tonic heat data
  • Yasser (BSc.): The use of machine learning tools for pain recognition

  • Aydogan (BSc.): Development of data acquisition and annotation system for enabling machine learning and affective computing
  • Blenk (MSc.): Echtzeitfähige Produktionssteuerung basierend auf einem Multi-Agenten-System mit Reinforcement Learning
  • Caliskan (MSc.): A Simple and Intuitive Way of Facial Animation
  • Englert (MSc.): Contour detection in recurrent center-surround networks
  • Fischer (MSc.): Neuer Ansatz für Nelson Regeln in der Prozesskontrolle mit Anwendungen im maschinellen Lernen
  • Franke (BSc.): Implementation and Evaluation oft he Minimal Complexity Machine
  • Füßinger (BSc.): Ein informationstheoretischer Regulafisierer für Knstliche Neuronale Netze in wechselnder Umgebung
  • Heyne (MSc.): Automatic Root Cause Analysis in Pre-Silicon Verification of Complex Superscalar Microprozessors with Machine Learning
  • Hofherr (MSc.): Indoor Localization by Combined Smartphone Sensor Data using Neural Networks
  • Kalischek (MSc.): Deep Domain Adaptation for facial expression analysis
  • Kneist (MSc.): Neural Networks for Web Page Segmentation
  • Landgraf (MSc.): Instance segmentation in bin-picking scenarios using convolutional neural networks
  • Lee (MSc.): Design and analysis of an emotional dataset consisted of biosignals, facial expressions and continuous self-annotations
  • Mehlhase (MSc.): Image matching with CNNs for medical images leveraging self-supervised learning and synthetic data
  • Müller (MSc.): Joint unsupervised learning of scene structure and motion
  • Oppel (Msc.): Active learning in affective computing utilizing deep neural networks
  • Park (MSc.): Visual emotion recognition in the wild at multi-user and group levels
  • Sauter (MSc.): Modellierung und Simulaton der Prozesskräfte beim außenrundschleifen
  • Schell (MSc.): Sequentielle Anwendung eines Multi-Layer Perceptron Netzwerks als Rotorlagensensor in permanent erregten Synchromaschinen
  • Schromm (MSc.): Implementierung und Untersuchung einer Hand Detection mittels eines Capsule Neural Networks
  • Srinivasan (MSc.): Deep Learning Algorithms for Emotion Recognition on Low Power Single Board Computers
  • Triep (MSc.): Learning to Manipulate a Robotic Arm Platform through Imitation Learning
  • Wagner (MSc.): Classifier-specific Feature Selection and Multiclass Classification of Road Users
  • Wurzberger (BSc.): Learning in Deep Radial Basis Function Networks
  • Yang (MSc.): Multi-Agent Actor-Critic Reinforcement Learning for Argumentative Dialogue Systems

  • Bauernschläfer (BSc.): Convolutional Neural Networks in der Mimikanalyse für die Schätzung von Schmerzniveaus
  • Chowdhury (MSc.): Deep Learning Architectures for RGB-D Sensor Fusion
  • El Sayegh (BSc.): Event-based motion computation for visual navigation in robotic vision
  • Gödecke (MSc.): Visualization and Exploratory Analysis of non-invasive Measurements in an Auditory Learning
  • Gloger (MSc.): The influence of error attribution on sensorimotor learning
  • Graml (MSc.): Learning Probabilistic Motion Primitives with Mixture-of Experts Architectures
  • Gruhler (MSc.): Konzeption und Implementierung eines neuronalen Netzes zur Gestenerkennung anhand von Millimeterwellen-Radarwellen
  • Horn (MSc.): Motion Classification and Height Estimation of Pedestrians Using Sparse Radar Date
  • Hurst (MSc.): Erkennung subjektiver Schmerzereignisse mittels Cross Corpus Analyse und Transferlernverfahren
  • Kaiser (MSc.): Machine Learning Classifiers in the Analysis of Button press Dynamics and fMRI Data in an Auditory Learning Experiment
  • Kraft (MSc.): Maschinelles Lernen zur Prognose der Prozesszeit von Änderungsvorhaben in der Fahrzeugentwicklung
  • Kraus (MSc.): Ship Classification Based on AIS Data with Machine Learning Methods
  • Kim (MSc.): End-to-End time-continuous emotion recognition for spontaneous interactions
  • Liebl (MSc.): Non-Contact Vital Signs Monitoring using ANN and Millimeter-Wave Radar Technology
  • Lutz (MSc.): Anwendung von Methoden der SA zur Analyse von E-Learning Kursen – Demonstration an einem realen Datensatz
  • Mahankali (MSc.): Stockwell-Transform using Neural Networks for Speaker Identification
  • Pirkani (MSc.): Analysis of Delays in Continuous Annotations and Physiological Signals for Affective State Characterization using Machine Learning Algorithms
  • Rimmel (BSc.): Strategielernen mit tiefen neuronalen Netzen
  • Schmid (MSc.): A mathematical approach to nonlinear support vector machines and approximate kernels with near random Gaussian Matrices
  • Sellner (MSc.): Distinguishing Pain from Emotion Using Facial Expressions
  • Tawik (BSc.): Optical Character Recognition for Arabic Letters
  • Thomas (MSc.): Transfer Learning Using Artificial Neural Networks for the Design of Personalized Models for Affect Recognition
  • Welz (MSc.): Extracting Semantic Radar Feature Grids using Recurrent Deep Convolutional Neural Networks
  • Xyu (MSc.): Rekurrente neuronale Netze zur Schmerzerkennung aus biophysiologischen Daten
  • Yousry (BSc.): Fall detection using event-based data
  • Yu (MSc.): Multi-Label Learning for Recognition in Emotions and Pain
  • Mamontov (MSc.): Optimization of recurrent neural networks sructure ba means of self-configuring genetic programming algotithm
  • Polonskaya (MSc.): Designing of feed forward neural networks architecture by means of self-configuring genetic programming algorithm
  • Skorokhod (MSc.): Reducing of dimension by participial swarm optimization

  • Akella (MSc.): Estimation of the user’s cognitive load utilizing audio-based deep neural network architectures and learning strategies
  • Amirian (MSc.): Compressed sensing in pattern recognition and machine learning for human computer interaction
  • Brunel (MSc.): Personenübergreifender Wissenstransfer zur Emotionserkennung
  • Bühler (BSc.): Schätzung von Ausfallzeitpunkten aus multimodalen Sensordaten mit Hilfe künstlicher neuronaler Netze
  • Engel (MSc.): Multi-modal myocontrol for bimanual manipulation
  • Frings (MSc.): Multimodale Emotionserkennung mit Autoencodern optimiert durch evolutionäre Algorithmen
  • Hartwig (MSc.): Deriving parallax environment maps from panorama cameras
  • Ismail (MSc.): Incorporating Modulatory Feedback in Convolutional Neural Networks
  • König (MSc.): Deep Lerning for Low Resolution Person Detection in Multispectral Videos
  • Lüke (Dipl.): Implementation and Simulation of Boolean Models for Gene Regulation on a Hybrid x86/FPGA System
  • Mundus (Dipl.): Support Vektor Maschinen für Mehrklassenprobleme mit Fuzzy-Lehrersignalen
  • Mustafa (BSc.): Interest point detecton and tracking for detecting fall events in video
  • Shymbarova (MSc.): Neuromorphe Implementierung eines kortikalen Kolumnenmodells
  • Steinacher (MSc.): Semi-Supervised-Learning-Verfahren zur Kategorisierung von Kundendatensätzen
  • Vincenot-Gassmann (MSc.): Maschinelle Lernverfahren zur Erkennung kognitiver Belastung anhand mimischer Expression

  • Bellmann (MSc): Diversität in Mehrklassifikatorsystemen
  • Cheng (MSc): Iterative Optimierung von Fuzzy Support Vektor Maschinen
  • Gloger(Dipl.): Formalizing visual representatins of spatial data using symbolic regression
  • Held (MSc): Bimodale Emotionserkennung aus gesprochener Sprache und Biosignalen
  • Heyne (BSc.): Structure-based image fusion
  • Hihn (MSc.): Multiple classifier systems for video-based affect recognition
  • Kindsvater (MSc.): Fusionsarchitekturen für die multimodale Affekterkennung unter Realzeitbedingungen
  • Krammer (MSc.): Adaptive und lernende Regler für Montageanwendungen mit Industrierobotern
  • Krebs (MSc.): Online vision-based multi-object tracking for vulnerable Rcad users
  • Kreiser (MSc.): Shadow Detection Based Surface Reconstruction
  • Löhr (MSc.): Objektsegmentierung und-erkennung auf Basis hyrider Punktwolken-Panoramabild Repräsentationen in einem industriellen Anwendungsfeld
  • Rottach (BSc.): Comparison between scene flow and dense stereo
  • Speck (MSc.): Entwicklung multimodaler Klassifikatoren durch Instanzenselektion
  • Wieluch (MSc.): Multimodal active learning in affective computing
  • Winterling (MSc.): Semi-supervised classification of radar based grid maps using deep learning methods