Student research projects & theses

The staff of the KBS will be pleased to advise you with current topic suggestions for your Bachelor/Master thesis.

Attention: Bachelor theses can also be carried out in pairs as long as the individual contribution can be assigned. For correspondingly more extensive tasks, please also see the Master's theses.

BACHELOR AND MASTER THESES

  • Research Internship Opportunity at Huawei Munich

    Huawei R&D Munich is looking for keen and motivated students to join us as research interns for the upcoming summer. This is a good chance for Master/PhD students to gain hands-on experience in LLM & Reinforcement Learning and to contribute to some exciting projects.

    Here are the details:

    • Position: Research Intern for LLM Fine-Tuning
    • Duration: 6 months (with possibility of extension)
    • Start Date: Summer 2024
    • Location: Munich
    • Requirements: currently enrolled in a Master / PhD program, with some background and interests in LLM & RL.

    Application Deadline: 20.6.2024

    LinkedIn posting: https://www.linkedin.com/jobs/view/3934264485

     

  • Optimizing Robot Interaction using Image-based Tactile Sensors (Wadhah Zai El Amri)

    Tactile sensing presents a promising opportunity for enhancing the interaction capabilities of today’s robots. DIGIT [1] is a commonly used tactile sensor that enables robots to perceive and respond to physical tactile stimuli.  

    In situations where the sensor is unavailable or experiment repetitions are costly, the value of a reliable, real-time simulation becomes evident. Such a simulation can effectively estimate sensor outputs for various touch scenarios. Such simulation would offer a good alternative to gathering data in different setups and environments. 

    While several studies have introduced simulations for the DIGIT sensor, such as TACTO [2] and Taxim [3], they predominantly rely on rigid-body simulations and overlook the crucial soft-body aspect of the sensor's gel tip. This oversight diminishes the accuracy of the simulations and their ability to faithfully replicate real-world tactile interactions. 


    Goals of the thesis

    • Collect real image outputs and simulated sensor surface deformation.
    • Train a machine learning algorithm to generate output images using DIGIT surface deformation mesh.
    • Perform manipulation/grasping tasks with a real UR5 robot to assess and evaluate the performance of the trained algorithm.


    Prior Knowledge or interest

    • Machine Learning
    • Robotic Operating System (ROS)
    • Python
    • Linux

    Related Work:

    • [1]: M. Lambeta, P.-W. Chou, S. Tian, B. Yang, B. Maloon, V. R. Most, D. Stroud, R. Santos, A. Byagowi, G. Kammerer, D. Jayaraman, and R. Calandra, “DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor With Application to In-Hand Manipulation,” IEEE Robotics and Automation Letters, vol. 5, no. 3, pp. 3838–3845, 2020.
    • [2]: S. Wang, M. Lambeta, P.-W. Chou, and R. Calandra, “TACTO: A Fast, Flexible, and Open-Source Simulator for High-Resolution Vision-Based Tactile Sensors,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3930–3937, 2022.
    • [3]: Z. Si and W. Yuan, “Taxim: An example-based simulation model for gelsight tactile sensors,” IEEE Robotics and Automation Letters, 2022.


    Contact


    We look forward to receiving your complete informative application documents.

    For further information and application, please contact Wadhah Zai El Amri at wadhah.zai@l3s.de 

  • Multi-objective Optimization for Robotic Gripper Design (Nicolas Navarro-Guerrero)

    The human hand is often used as the gold standard or goal for robotic manipulation, and haptic perception and illustration are often used to illustrate sophisticated robots and AIs. However, despite our fascination for anthropomorphic robotic hands, the trend in industrial applications and robotic challenges is to use simpler designs consisting of parallel grippers or suction cups. This trend is not necessarily only due to the complexity of matching the human hand's dexterity, robustness and perceptual capabilities but also to the reality that anthropomorphic hands may not be required to achieve the human skill level. For instance, the winner of the Amazon Picking Challenge used an end effector based on a suction system. In the DARPA Robotics Challenge, 15 of 25 teams used an underactuated hand with three or four fingers (Piazza et al., 2019), while none of the remaining ten teams used a fully actuated anthropomorphic hand (Piazza et al., 2019). Even in the Cybathlon, the winner of the Powered Arm Prosthesis Race used a body-powered hook (Piazza et al., 2019).

    In addition, the synergistic combination of all three subsystems, including the mechanical aspects, perception, and control, might be more critical than an anthropomorphic robotic hand to match and potentially surpass human dexterous manipulation capabilities.

    All of which prompts the question of whether we need anthropomorphic robotic hands. This thesis aims to answer this question from a multi-objective optimization point of view.


    Goals of the thesis

    • Implement a Genetic Algorithm or another multi-objective optimization algorithm to develop a general-purpose robotic manipulator. The manipulator could be based on the human hand, ideally reducing complexity while keeping most of its dexterity.


    Prior Knowledge or interest

    • Machine Learning
    • Python
    • Linux

    Related Work:

    • Cheney, N., MacCurdy, R., Clune, J., & Lipson, H. (2013). Unshackling Evolution: Evolving Soft Robots with Multiple Materials and a Powerful Generative Encoding. Annual Conference on Genetic and Evolutionary Computation (GECCO), 167–174. https://doi.org/10.1145/2463372.2463404
    • Coevoet, E., Morales-Bieze, T., Largilliere, F., Zhang, Z., Thieffry, M., Sanz-Lopez, M., Carrez, B., Marchal, D., Goury, O., Dequidt, J., & Duriez, C. (2017). Software Toolkit for Modeling, Simulation, and Control of Soft Robots. Advanced Robotics, 31(22), 1208–1224. https://doi.org/10.1080/01691864.2017.1395362
    • Faure, F., Duriez, C., Delingette, H., Allard, J., Gilles, B., Marchesseau, S., Talbot, H., Courtecuisse, H., Bousquet, G., Peterlik, I., & Cotin, S. (2012). SOFA: A Multi-Model Framework for Interactive Physical Simulation. In Y. Payan (Ed.), Soft Tissue Biomechanical Modeling for Computer Assisted Surgery (pp. 283–321). Springer. https://doi.org/10.1007/8415_2012_125
    • Piazza, C., Grioli, G., Catalano, M. G., & Bicchi, A. (2019). A Century of Robotic Hands. Annual Review of Control, Robotics, and Autonomous Systems, 2(1), 1–32. doi.org/10.1146/annurev-control-060117-105003


    Contact


    We look forward to receiving your complete informative application documents.
    For further information and application, please contact Nicolas Navarro-Guerrero at nicolas.navarro@l3s.de

  • Forecasting Product Demand (Hubert Truchan)

    I would like to announce that we are now accepting applications for a Master's thesis, Bachelor's thesis on the topic of "Time series forecasting with the use of neural networks." This is an exciting and innovative area of study that has the potential to make significant contributions to the field of computer science.

    In this thesis, students will have the opportunity to explore the use of neural networks for a time series forecasting problem that has traditionally been difficult to solve using traditional machine learning techniques. The goal of this research is to develop new algorithms and techniques that can accurately forecast time series data, with a focus on improving the performance and accuracy of existing methods.

    Students interested in this topic should have a strong background in computer science, with a particular focus on machine learning and neural networks. Knowledge of time series analysis and regression techniques is also helpful but not required.

    If you are interested in pursuing this topic for your Master's thesis, Bachelor's thesis, please submit your application via email [truchan@L3S.DE] with subject [Thesis, Forecasting Product Demand, Your Name]. We look forward to reviewing your application and discussing this exciting opportunity with you.

    Possible start: 15.01.2024

  • Neural Architecture Fusion of Image and Time Series Data (Hubert Truchan)

    The world of neural networks, especially in the areas of image classification and time series processing, continues to expand rapidly. In this master's thesis project, we aim to explore the fusion of these two domains by experimenting with innovative combinations of deep learning,  neural network architectures.

    Students who are interested in delving deep into neural network architectures and unveiling efficient methods to merge them into a cohesive, high-performance system are encouraged to apply. 
    The learning materials are provided and you will have the opportunity to implement your own ideas.

    The thesis steps include:

    1. A comprehensive literature review on the topic
    2. Undertaking exploratory data analysis of both images and time series
    3. Engaging in feature engineering to enhance model performance
    4. Signals modelling to understand data patterns
    5. Evaluating model performance and interpreting results
    6. Refining the model to achieve optimal outcomes
    7. Documenting the entire process and creating a final report

    Qualifications:

    • Strong Python programming skills
    • Solid theoretical understanding of deep learning
    • Experience with neural models

    How to Apply:
    If you want to find out more about the topic or have your own propositions,  please send an email to [truchan@L3S.DE] with the subject: "Master's thesis, Neural Architecture Fusion of Image and Time Series Data, [Your Name]"

    Possible project start: 01.11.2023

  • Exploring ChatGPT for Drone Applications: Prompts, Dialogues, and Task Adaptability (Dr. Marco Fisichella)

    This thesis aims to explore the application of OpenAI's ChatGPT in the domain of drones. The objective is to investigate the feasibility and effectiveness of leveraging ChatGPT for various drone tasks in both real and simulated environments. The thesis proposes a comprehensive strategy that combines prompt engineering principles and software engineering skills to facilitate ChatGPT's adaptation to different drone tasks.

    The study will evaluate the efficacy of different prompt engineering techniques, dialog strategies, and software architectural decisions in the context of executing drone-related tasks. Special emphasis will be placed on ChatGPT's capabilities in utilizing free-form dialogue, code synthesis, task-specific prompting functions, and closed-loop reasoning through dialogues.

    The thesis will cover a wide range of drone tasks, including aerial navigation, object detection, and path planning. The primary objective is to ascertain whether ChatGPT can effectively solve these tasks while enabling users to interact predominantly through natural language instructions.

    Throughout the research, both real-world drone experiments and simulated scenarios will be conducted to assess the performance and adaptability of ChatGPT. Software engineering best practices will be employed to ensure the robustness and scalability of the system.

    The anticipated outcome of this thesis is to demonstrate the potential of ChatGPT as a valuable tool for drone applications, showcasing its ability to enhance user interactions and task execution through natural language instructions.

    Keywords: ChatGPT, drones, prompt engineering, natural language interaction, aerial navigation, object detection, path planning, simulated environment, software architecture.


    Contact
    We look forward to receiving your complete informative application documents.
    For further information and application, please contact Dr. Marco Fisichella at mfisichella@L3S.de

  • Implementation of Low Bandwidth and Latency Image Transport for ROS(2) (Nicolas Navarro)

    Although camera technology has developed considerably in the consumer
    market, most robots still use inexpensive webcams (or RGB-D cameras),
    which cannot deal with challenging lighting conditions and lack optical
    image stabilization, which significantly impacts the quality of the
    images generated. Additionally, image and videos "transport" packages
    (e.g. https://wiki.ros.org/image_transport_plugins) use very dated
    technologies with a detrimental effect on band-wide usage and latency in
    robot applications.

    Possible goals of the thesis
    * Implement an Open Source ROS(2) package with improved image compression
    * Implement an Open Source ROS(2) package with high-dynamic range (HDR)
    capabilities

    The user should be able to enable one or both at a time. The compression
    and HDR levels should be adjusted from 0 to a maximum to be defined.

    Prior Knowledge or interest
    * Machine Learning
    * Python
    * Linux

    Related Work:
    One version of the thesis could be to improve the high dynamic range
    (HDR) in robotics applications.
    > Benzi, M., Escobar, M.-J., & Kornprobst, P. (2018). A Bio-Inspired
    Synergistic Virtual Retina Model for Tone Mapping. Computer Vision and
    Image Understanding, 168, 21–36. https://doi.org/10.1016/j.cviu.2017.11.013

    And another to compress images in order to lower bandwidth requirements:
    See Section 1.3.3.
    > Escobar, M.-J., Alexandre, F., Viéville, T., & Palacios, A. G. (2021).
    Rapid Prototyping for Bio–Inspired Robots. In Rapid Roboting: Recent
    Advances on 3D Printers and Robotics (1st ed.). Springer International
    Publishing. https://www.springer.com/gp/book/9783319400013

    Contact
    We look forward to receiving your complete informative application
    documents. For further information and application, please contact
    Nicolas Navarro-Guerrero at nicolas.navarro@l3s.de

  • Analysis of Propagation of Vibrations and Body-Borne Sound in a Robotic Hand (Nicolas Navarro)

    Haptic sensors span a broad range of technologies. The main focus of the
    sensors is to increase the recognition accuracy of both textures and the
    location of contact points. However, these sensors are mechanically
    fragile and mounted externally to robotic systems to increase accuracy,
    limiting the use of those sensors to applications that are kind to the
    sensors. For use in harsh applications or complementary to those
    existing sensors, this project aims to develop a machine
    learning-oriented solution capable of using body-borne vibrations to
    classify objects' texture and location of haptic interaction. This
    strategy allows mounting the sensors inside the robot, protected from
    external perturbance. Although this technology is not as accurate as
    other technologies, it promises to enable a degree of haptic perception
    anywhere the robot's outer shell (and electronics) can withstand.
    The technology has been validated in applications of multimodal object
    recognition, e.g., by Bonner et al. 2021 and Toprak et al. 2018. Some of
    the following steps include the development of the algorithms to perform
    localization of multiple points of contact between the robot and
    external objects.

    Goal
    * Systematic collect sound and vibration data from a robotic hand
    * Determine ideal placement of sensors
    * Perform sound-source localization of the source of the vibration or
    sound within the robotic hand

    Prior Knowledge or interest
    * Machine Learning
    * Python
    * A electronics or mechatronics background


    Related Work:
    > Navarro-Guerrero, N., Toprak, S., Josifovski, J., & Jamone, L.
    (2023). Visuo-Haptic Object Perception for Robots: An Overview.
    Autonomous Robots, 27.
    https://link.springer.com/article/10.1007/s10514-023-10091-y

    > Bonner, L. E. R., Buhl, D. D., Kristensen, K., & Navarro-Guerrero, N.
    (2021). AU Dataset for Visuo-Haptic Object Recognition for Robots.
    figshare. https://doi.org/10.6084/m9.figshare.14222486

    > Toprak, S., Navarro-Guerrero, N., & Wermter, S. (2018). Evaluating
    Integration Strategies for Visuo-Haptic Object Recognition. Cognitive
    Computation, 10(3), 408–425. https://doi.org/10.1007/s12559-017-9536-7


    We look forward to receiving your complete informative application
    documents. For further information and application, please contact
    Nicolas Navarro-Guerrero at nicolas.navarro@l3s.de

  • Development of 3D printed fingerprint for object recognition (Nicolas Navarro)

    Multimodal object recognition is still an emerging and active field of
    research. Haptic sensors span a broad range of technologies. The main
    focus of the sensors is to increase the recognition accuracy of both
    textures and the location of contact points. However, these sensors are
    mechanically fragile and mounted externally to robotic systems to
    increase accuracy, limiting the use of those sensors to applications
    that are kind to the sensors. For use in harsh applications or
    complementary to those existing sensors, this project aims to develop a
    machine learning oriented solution capable of using body-borne
    vibrations to classify objects' texture and location of haptic
    interaction. This strategy allows mounting the sensors inside the robot,
    protected from external perturbance. Although this technology is not as
    accurate as other technologies, it promises to enable a degree of haptic
    perception anywhere the robot outer shell (and electronics) can withstand.
    The technology has been validated in applications of multimodal object
    recognition, e.g., by Bonner et al. 2021 and Toprak et al. 2018. Some of
    the following steps include 1) the development of "robotic fingerprints"
    that maximize the body-borne vibrations, 2) and later developing the
    algorithms to perform localization of multiple points of contact between
    the robot and external objects.

    Goal
    * development of 3D printed robotic fingerprints for robotic grippers.
    - systematic optimization of fingerprint's texture
    - systematic selection of material for different applications,
    underwater manipulation, humanoid robots, etc.
    - creation of a haptic dataset

    Prior Knowledge or interest
    * CAD
    * 3D printing
    * Rapid prototyping

    Related Work:
    > Navarro-Guerrero, N., Toprak, S., Josifovski, J., & Jamone, L.
    (2023). Visuo-Haptic Object Perception for Robots: An Overview.
    Autonomous Robots, 27.
    https://link.springer.com/article/10.1007/s10514-023-10091-y

    > Bonner, L. E. R., Buhl, D. D., Kristensen, K., & Navarro-Guerrero, N.
    (2021). AU Dataset for Visuo-Haptic Object Recognition for Robots.
    figshare. https://doi.org/10.6084/m9.figshare.14222486

    > Toprak, S., Navarro-Guerrero, N., & Wermter, S. (2018). Evaluating
    Integration Strategies for Visuo-Haptic Object Recognition. Cognitive
    Computation, 10(3), 408–425. https://doi.org/10.1007/s12559-017-9536-7


    We look forward to receiving your complete informative application
    documents. For further information and application, please contact
    Nicolas Navarro at nicolas.navarro@l3s.de

  • Quantification of the effect of feedback in Interactive Reinforcement Learning (Nicolas Navarro)

    As shown in multiple research, providing feedback to autonomous learning agents can speed up learning. However, the quantification/characterization of different aspects of feedback, such as feedback quantity, quality, temporal and spatial misalignments, etc., in learning speed, performance, and other relevant metrics is still an open question. This question not only addresses theoretical aspects of the learning algorithms but is also very relevant for application in real systems because although feedback might be beneficial, (human-)feedback is also expensive and adds complexity to the systems. Thus, it is essential to know the minimal requirements for the (human-)feedback to achieve a significant increment in performance, learning speed etc., that it is worth the added complexity. Hence, this project presents a series of questions that can be addressed independently to achieve a deeper understanding of the role of feedback in a learning system's performance.

    Several assumptions and simplifications can be made to facilitate the study of these questions. These include the use of binary and low-dimensional feedback, simulated environments, and the use of autonomous teachers. Moreover, this project will be studied in a robot reaching task for a KUKA LBR iiwa, a robotic arm with 7 degrees of freedom (DoF). Configurations from 1 to 7 DoF will be used to study feedback effects at different levels of task complexity.

    This project will use artificial feedback and primarily be studied in simulated environments. Eventually, once a better understanding of the effects of feedback is obtained, experiments with real users will be carried out. Several thesis directions are possible, which will be discussed with the candidates. These include:

    - Quantifying the Effect of Feedback Accuracy in IRL Performance
    - Quantifying the Effect of Feedback Quantity in IRL Performance (e.g., binary, scalar value, or vector)
    - Quantifying the Effect of Feedback Budget in Interactive Reinforcement Learning (e.g., early, uniform, late)
    - Quantifying the Effect of Time-Delayed Feedback in IRL Performance
    - Policy Shaping in IRL for Dynamical System using Binary and Other Low-Dimensional Feedback


    Prior Knowledge or interest in
    - Reinforcement Learning and Machine Learning
    - Human-Robot Interaction
    - Python, Latex, git, Linux


    Related Work:
    Harnack, D., Pivin-Bachler, J., & Navarro-Guerrero, N. (2022). Quantifying the Effect of Feedback Frequency in Interactive Reinforcement Learning for Robotic Tasks. Neural Computing and Applications. Special Issue on Human-aligned Reinforcement Learning for Autonomous Agents and Robots. https://doi.org/10.1007/s00521-022-07949-0

    Stahlhut, C., Navarro-Guerrero, N., Weber, C., & Wermter, S. (2015). Interaction in Reinforcement Learning Reduces the Need for Finely Tuned Hyperparameters in Complex Tasks. Kognitive Systeme, 3(2). https://doi.org/10.17185/duepublico/40718


    We look forward to receiving your complete informative application documents. For further information and application, please contact Nicolás Navarro-Guerrero.

    Contact:
    Nicolas Navarro  at nicolas.navarro@l3s.de

  • Continual Learning for Affective Robotics (Nicolas Navarro)

    Operating in human-centred environments, social and affective robots need to actively participate in the human ‘affective loop’, sensing and interpreting human socio-emotional behaviours while also learning to respond in a manner that fosters their social and emotional wellbeing. Current Machine Learning (ML)-based solutions for realising affective capabilities in robots, be it robust affect perception or behaviour generation, are primed towards the generalisation of application. Affective robots, on the other hand, need personalised interaction capabilities that, sensitive to an individual’s socio-emotional behaviour, can adapt affective interactions towards them, expanding their learning on the go to include novel information, while ensuring past knowledge is preserved. This talk will explore the Continual Learning (CL) paradigm for enabling such adaptation capabilities in robots at every stage of the human ‘affective loop’.

    Laying out the foundational formulations that translate key CL principles for affective learning, it will reflect upon the key desiderata for affective robots in terms of continual and personalised affect perception, and context-appropriate behaviour generation.

    Personalised affect perception may allow for incremental learning of expression classes while being sensitive to individual differences in expression. Furthermore, it is also important to guard against biases arising from imbalances in data distributions, especially with respect to demographic attributes of gender (male/female, in this case) and race, that may negatively impact sensitive populations. This talk will explore CL-based learning to tackle these challenges, highlighting the challenges as well as benefits of adopting CL to achieve fairer affect recognition. Finally, embedding such personalised learning capabilities in a robotic wellbeing coach, a practical framework will be presented that can allow social and affective robots to dynamically adapt human-robot interactions by generating naturalistic responses, sensitive to the participants’ affective state. Overall, the theoretical formulations and practical frameworks discussed in this talk will aim to initiate a novel field of enquiry exploring the benefits of CL-based learning for affective robots.

    Short Bio:
    Nikhil Churamani is currently a Post-Doctoral Research Assistant at the Affective Intelligence and Robotics (AFAR) Laboratory of the Department of Computer Science and Technology, University of Cambridge. He completed his PhD on ``Continual Learning for Affective Robotics'' at the AFAR Lab, Cambridge where he was funded by the EPSRC International Doctoral Scholarship and the Premium Research Studentship of the department. His research interests include Affective Computing, Continual Learning, Computer Vision, Deep Learning and Human-Robot Interaction. His current research focuses on Continual Learning for Affective Robotics investigating Continual Lifelong Learning of Affect for social robots, focused on Human-Robot Interaction and affect-driven learning.
    Website:  https://www.cl.cam.ac.uk/~nc528

    Contact:
    Nicolas Navarro  at nicolas.navarro@l3s.de

  • An Interpretable Federated Multivariate Time Series Classification method (Raneen Younis, Dr. Zahra Ahmadi, Dr. Marco Fisichella)

    Increasing privacy concerns have led to decentralized and federated machine learning techniques that allow individual clients to consult and train models collaboratively without sharing private information. Some of these applications, such as in medicine and healthcare, require that final decisions be interpretable. A common form of data in these applications is multivariate time series. While several approaches have addressed the interpretability of Deep Learning models for multivariate time series data in a centralized environment, less effort has been made in a federated setting.

    Previous work has presented FLAMES2Graph, a horizontal federated learning framework designed to interpret Deep Learning decisions made by individual clients. FLAMES2Graph extracts and visualizes the subsequences of the inputs that are highly activated by a convolutional neural network. An evolution graph is also created to capture the temporal dependencies between the extracted subsequences. The federated learning clients only share this temporal evolution graph with the central server instead of training the model weights to build a global evolution graph.

    FLAMES2Graph assumes that all clients have a fixed number of nodes when constructing the graph. However, this assumption can lead to the loss of important information from other clients. Therefore, a new approach had to be developed to improve the construction of the graph using a dynamic graph. This led to challenges such as defining node similarity, merging the dynamic graphs, and updating the clients' local models.

    Please reach out if you're interested in pursuing this research area for your master's thesis or if you have any additional inquiries. Contact: Raneen Younis at younis@l3s.de

  • Neural Architecture Fusion for Machine State Recognition (Hubert Truchan)

    We are excited to announce a new Master's thesis opportunity in the area of image classification and time series processing. The topic of the thesis is "The Neural Architecture Fusion for Machine State Recognition."

    This thesis project aims to explore and experiment with novel combinations of convolutional neural network (CNN) architectures for improved machine state recognition. Students who are interested in diving deep into CNN architectures and discovering efficient methods to fuse them into a single, highly accurate system are encouraged to apply.

    Thesis steps include:

    1. Literature review
    2. Exploratory Data Analysis (images and time series)
    3. Feature engineering
    4. Signals modeling
    5. Model evaluation and results interpretation
    6. Model refinement
    7. Creating the documentation and final report

    Required qualifications for this thesis project:

    • Strong Python programming skills
    • Solid theoretical understanding of deep learning
    • Experience with convolutional neural models

     

    If you are interested in this topic, please send an email to [truchan@L3S.DE] with the subject: "Master's thesis, Neural Architecture Fusion for Machine State Recognition, Your Name".

  • Robustness of Stable Diffusion models (Sandipan Sikdar)
  • Predictive Quality Control of Geared Motors (Hubert Truchan)

    We are now accepting applications for a Master's thesis on the topic of "Image Classification with modern neural models." This is an innovative and cutting-edge area of study that has the potential to make significant contributions to the field of computer science.

    In this thesis, students will have the opportunity to explore the use of modern neural models for image classification, which has recently made impressive progress and can be applied in many domains. The goal of this research is to develop new algorithms and techniques that can accurately classify images, with a focus on improving the performance and accuracy of existing methods.

    Students interested in this topic should have a strong background in computer science, with a particular focus on machine learning and neural networks. Knowledge of image processing and classification techniques is also helpful but not required.

    Required qualifications for this thesis project:

    • Strong Python programming skills
    • Solid theoretical understanding of deep learning
    • Experience with convolutional neural models

    If you are interested in this topic, please send an email to [truchan@L3S.DE] with the subject: "Master's thesis, Predictive Quality Control of Geared Motors, Your Name".

    Hubert Truchan

  • Digitale Transformation in der Medizin (Prof. Nejdl)

    Details zum Projekt

    Über das Programm:

    • ein strukturiertes Ausbildungsprogramm zur Förderung der Kooperation zwischen Medizin und Informatik
    • Gemeinsame Abschlussarbeiten für:
      • Promovierende der Humanmedizin (12 Monate)
      • Masterstudierende der Informatik (6 Monate)
        … im Themenbereich Digitale Transformation in der Medizin

    Benefits:

    • Eine monatliche finanzielle Unterstützung und eine enge wissenschaftliche Betreuung werden gestellt,
    • Projektveranstaltungen im Plenum und Lehrveranstaltungen bilden das Rahmenprogramm.

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Wolfgang Nejdl

  • Multisensory neural models (Hubert Truchan)

    Healthcare and production systems must be constantly monitored due to frequent changes in the operating environment, machine parameters and subject state. Medicine and production processes impose strict requirements for precisely monitoring machine parameters therefore, online feedback and high precision of the predicted process parameters are needed. Furthermore, rapid digitalization and advances in sensor systems enable direct and indirect measurements of complex processes. Thus, the task of the master thesis is to use the advantage of the multiple sensors systems to precisely map the underlying operating state.

    Required qualifications for all projects/thesis:

    • Strong Python programming skills
    • Solid theoretical understanding of deep learning
    • Experience with convolutional network models

    If you are interested in this topic, please send an email to [truchan@L3S.DE] with the subject: "Master's thesis, Multisensory neural models, Your Name".

    Leitung und Ansprechpartner der Abschlussarbeit

    Hubert Truchan

  • Trustworthy Epidemic Related Information Extraction From Microblogs

    Microblogging platforms like Twitter provide rapid access to vital information during epidemic outbreaks. Affected communities come to these sites to report, seek or share details such as disease symptoms, transmission, and prevention measures. Besides, the massive data is highly helpful for health organizations to obtain situational updates, and plan for treatment/prevention measures. However, disease-related messages are immersed in a high volume of irrelevant information. It thus requires efficient methods to handle data overload and prioritize various types of information. In this work, we aim at building a trustworthy machine learning model to detect disease-related tweets and classify the data into granularity levels such as symptoms, transmission, prevention, etc. The predictive model should be both highly accurate and interpretable. 

    Required qualifications for all projects/thesis:

    • Strong programming background in Python
    • Good knowledge of Machine Learning and Deep Learning

    Leitung und Ansprechpartner der Abschlussarbeit

    Marco Fisichella, Koustav Rudra, Thi Huyen Nguyen

  • Graph Neural network: Sensing On-Street Parking Space Availability (Dr. Marco Fisichella)

    Monitoring the occupancy of on-street parking spaces on a city-wide scale is still an open issue. Past research demonstrated the viability of parking crowd-sensing by means of standard on-board sensors of probe vehicles, foreseeing the use of high-mileage vehicles, like taxis. Nevertheless, the achievable spatio-temporal sensing coverage has never been deeply investigated.

    In this master thesis we will investigate the suitability of taxi fleets combined to parking sensors to crowd-sense on-street parking availability via graph neural network.

    An ideal candidate should have:

    • Strong background in python and deep learning
    • Motivation behind learning and exploring the data

    Related Papers:

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Marco Fisichella

  • Inverse design of nanophotonic structures with deep neural networks

    Optical metasurfaces represent a revolutionary tool to manipulate the behaviour of light at the nanoscale. They can reduce the footprint of traditional optical components, e.g., flat lenses, and achieve optical properties otherwise not available (tunable beam steering, wavefront manipulation, etc), thus finding applications in all domains of optics and photonics. The design of metasurfaces rely on full-wave simulations combined with optimization algorithms, such as topology optimization, genetic algorithms, and particle swarm. In recent years, deep neural networks have been introduced for the inverse design of metasurfaces and nanostructured materials in general. The selected master’s student will work on the develop ment of deep neural network techniques for the inverse design of optical metasurfaces.

    Required qualifications for all projects/thesis:

    • Excellent programming skills in Python, and experience working with Pytorch, Keras, or TensorFlow.
    • Very good knowledge of electromagnetics theory
    • Excellent communication in English.

    Leitung und Ansprechpartner der Abschlussarbeit

    Marco Fisichella, Antonio Calalesina, Niloy Ganguly and Gregory Palmer

  • Erfassung des Calciumcyclings und Kontraktion von humanen Herzmuskelzellen

    Kontraktion und Generierung eines intrazellulären Calciumtransienten ist Grundlage der physiologischen Funktion der Herzmuskelzelle. Akkurate Messung der Veränderung dieser zellulären Funktion ist essentiell, um Pathomechanismen der Erkrankungen zu entschlüsseln und neue Therapien zu testen.

    Von humanen induzierten pluripotenten Stammzellen abgeleitete Kardiomyozyten (hiPSC-CMs) stellen ein vielversprechendes humanes in vitro Zellmodell dar, um die grundlegende kardiovaskuläre Physiologie zu studieren, Effekte der Mutationen zu untersuchen und neue Therapien zu erproben. Vor kurzem wurden Matlab-basierte Codes entwickelt, die Studien der Kontraktilität und Calcium-Homöostase in hiPSC-CMs in automatisierter Weise ermöglichen. Dabei ermöglicht "SarcTrack", ein Matlab-basierter Code, aus Video-Datensätzen mit Fluoreszenz-markierten Sarkomerstrukturen die Kontraktion der Zellen voll-automatisch zu analysieren. Damit sollen aus großen Datensätzen Kontraktionsparameter wie minimale und maximale Sarkomerlängen sowie Kontraktions- und Relaxationskinetik gewonnen werden. Dazu muss aber die Robustheit des Codes weiter erhöht werden.

    In unseren Projekten haben wir hiPSC-CMs mit fluoreszenten Sarkomeren bereits erfolgreich produziert und mittels modernster high-framerate Konfokalmikroskope Videosequenzen aufgenommen. Die Datensätze wurden bereits auf einem Hochleistungs-HPC-Cluster analysiert.

    In diesem Projekt sollen gezielt Parameter des SarcTrack Codes systematisch moduliert werden und der Output für die bereits vorhandenen Datensätze optimiert werden. Eine geeignete Umgebung soll dafür geschaffen werden. Darüber hinaus soll ein weiterer bereits entwickelter Code "CalTrack" an der MHH etabliert und für die synchrone Analyse der Kalziumsignale in den Videosequenzen genutzt werden.

    Was wir bieten:

    Wir bieten eine sorgfältige Einarbeitung und sehr gute Betreuung durch direkte Ansprechpartner/in sowie eine spannende und hoch-moderne translationale Forschung in einem interdisziplinären Team. Der Beginn der Arbeit ist nach Absprache sofort möglich.

    Was Sie mitbringen:

    1. Theoretische und praktische Erfahrung mit Matlab
    2. Eigenständige, zuverlässige Arbeitsweise und Teamfähigkeit
    3. Sorgfältige Dokumentationsfähigkeit, hohe Motivation

    PDF-Flyer

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. PhD Natalie Weber Dr. Andre Zeug

  • Developing Black-box Adversarial Attacks on Speech Emotion Recognition Models

    Nowadays, speech emotion recognition (SER) has been essential in human-computer interaction. The rapid development of deep learning for SER has become a popular research area. However, deep neural networks were shown vulnerable to external attacks, especially adversarial attacks, in which the adversarial examples are generated by adding human-indistinguishable perturbations to the original real samples. In SER, adversarial examples may lead to misclassification, resulting in invalid and misinterpreted interactions with users.

    Different from white-box adversarial attacks which require the data sources and targeted models' parameters, black-box adversarial attacks cannot obtain either the data source or parameters. Therefore, it is more challenging to generate black-box adversarial examples. The transferability of black-box adversarial attacks means the capability of an attack model on transferring among multiple targeted models. Enhancing the transferability can help save costs for learning a unified attacker than training multiple independent ones. Lifelong learning has been used to improve the transferability of black-box adversarial attacks. The goal of this topic is to further improve the black-box adversarial attacks' transferability.

    An ideal candidate should have:

    1. Good background knowledge of signal processing and deep learning.
    2. Strong programming skills in Python (Pytorch, Keras, or TensorFlow).

    Leitung und Ansprechpartner der Abschlussarbeit

    Zhao Ren

  • Computer Audition for Health Care: Integrating Machine Learning into Acoustic Analytics

    In real life, sound is an essential component of human perception of the world. Especially, human speech and human body sounds can reflect important physiological and psychological health information. Computer audition aims to teach a machine perceiving speech and audio signals by integrating signal processing, machine learning, and deep learning techniques. Recent advances in computer audition have proven promising in digital health applications, from disease diagnosis to therapies. Moreover, the rapid development of smartphones, tablets, and wearable devices promotes mobile health for contactless diagnosis and/or remote monitoring.

    Our goal is to develop robust and explainable deep neural networks for automated diagnosis from speech/audio signals. Possible topics include, but are not limited to:

    1. Developing COVID-19 detection models from multiple modalities, such as sounds (speech, cough, and breathing) and symptoms.
    2. Exploring the effect of preprocessing in automatic auscultation from phonocardiogram signals
    3. Respiratory disease diagnosis from sounds

    An ideal candidate should have:

    1. Good background knowledge of signal processing and deep learning.
    2. Strong programming skills in Python (Pytorch, Keras, or TensorFlow).

    Leitung und Ansprechpartner der Abschlussarbeit

    Zhao Ren

  • Counterfactual Explanation of Document Retrieval Models (Dr. Koustav Rudra)

    Document retrieval consists of two phases 1. First phase retrieval where a set of 100/1000 documents are retrieved for each query from a corpus of million documents, 2. In the second phase we deploy a ranker to rerank these documents based on their relevance to the query. This reranker is pretty complex and is based on deep neural models. The users don’t have any clue related to the decisions made by the ranker. Hence, in this task, our objective is to understand the reasoning behind the ranking process i.e., which words or phrases are relevant for the ranking, what are the changes we should make in the document to influence the ranking. We want to address following counterfactual explanation strategies over the ranking algorithms.

    1. For a given query, we rerank documents retrieved by first stage retrieval. If we take a pair of documents, how much change/perturbation do we have to make to those documents to change their ranking order. The changes to the documents should be minimal and topic of the document should not be changed drastically. To understand the important terms or phrases we can try different feature based approach such as gradients/ attention scores and make the changes in the documents.

    2. In some cases, web content writers want to understand the ranking strategies of search engines and use that information to create the content so that search engines promote their web pages in the top of the list. In case 1, we assume blackbox ranker model i.e., we cannot update or modify the model parameters. We are only allowed to modify the document content to some extent. However, in this scenario, we focus on a specific document and try to update our model parameters to promote that document i.e., the document should appear higher in the list.

     

    An ideal candidate should have:

    1. Strong background in python and deep learning
    2. motivation behind learning and exploring the data
    3. knowledge about basic IR concepts

     

    Related Papers:

    1. Measurable Counterfactual Local Explanations for Any Classifier

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Koustav Rudra

  • Question-Answering System for Legal Documents (Dr. Koustav Rudra)

    Question answering is a classical task in information retrieval. However, this might be a more complicated for the legal documents. Legal documents contain different sections such as fact, argument, judgement, etc. Hence, we may have to explore this section information along with the text to develop a better legal retrieval system. In some cases, answers are also not straightforward and complete. Hence, the retrieval system has to take care of all such complications.

     

    An ideal candidate should have:

    1. Strong background in python and deep learning
    2. motivation behind learning and exploring the data

     

    Related Papers:

    1. Dense passage retrieval for open-domain question answering

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Koustav Rudra

  • In-depth Analysis of Negative Sampling Strategies in Dense Retrieval (Dr. Koustav Rudra)

    Recent developments in representational learning for information retrieval can be organized in a conceptual framework that establishes two pairs of contrasts: sparse vs. dense representations and unsupervised vs. learned representations. Sparse learned representations can further be decomposed into expansion and term-weighting components. From Dense supervised retrieval scenario, we have methods like Dense Passage Retrieval (DPR), COIL, CLEAR. The performance of such dense supervised approaches are heavily dependent on the training procedure and how good the negative samples are. There are various strategies through which negative samples can be collected. For a given query, we can randomly sample negative documents from all the irrelevant ones. On the other hand, we can also sample negative documents that are close to the query i.e., having overlap with query terms but still irrelevant. The objective of this work is to understand the influence of different negative sampling strategies in the performance of different dense supervised retrieval set ups.

     

    An ideal candidate should have:

    1. Strong background in python and deep learning
    2. motivation behind learning and exploring the data
    3. knowledge about basic IR concepts

     

    Related Papers:

    1. Complementing lexical retrieval with semantic residual embedding 
    2. COIL : Revisit Exact Lexical Match in Information Retrieval with Contextualized Inverted List
    3. Distilling dense representations for ranking using tightly-coupled teachers

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Koustav Rudra

  • Precedent & Statute Retrieval Task in Legal Documents (Dr. Koustav Rudra)

    In countries following the Common Law system (e.g., UK, USA, Canada, Australia, India), there are two primary sources of law – Statutes (established laws) and Precedents (prior cases). Statutes deal with applying legal principles to a situation (facts / scenario / circumstances which lead to filing the case). Precedents or prior cases help a lawyer understand how the Court has dealt with similar scenarios in the past, and prepare the legal reasoning accordingly.

    When a lawyer is presented with a situation (that will potentially lead to filing of a case), it will be very beneficial to him/her if there is an automatic system that identifies a set of related prior cases involving similar situations as well as statutes/acts that can be most suited to the purpose in the given situation. Such a system shall not only help a lawyer but also benefit a common man, in a way of getting a preliminary understanding, even before he/she approaches a lawyer. It shall assist him/her in identifying where his/her legal problem fits, what legal actions he/she can proceed with (through statutes) and what were the outcomes of similar cases (through precedents).

     

    In this project our objective is to propose models for the following two tasks:

    • Task 1 : Identifying relevant prior cases for a given situation
    • Task 2 : Identifying most relevant statutes for a given situation

     

    An ideal candidate should have:

    1. Strong background in python and deep learning
    2. motivation behind learning and exploring the data

     

    Related Papers:

    1. Identification of Rhetorical Roles of Sentences in Indian Legal Judgments
    2. Overview of the FIRE 2019 AILA Track: Artificial Intelligence for Legal Assistance

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Koustav Rudra

  • Privacy Attacks on Graph Neural Networks (Dr. Megha Khosla)

    Machine learning (ML) algorithms have been applied on various applications including privacy-sensitive application such as in healthcare and finance.

    Similarly, many real world applications can be modeled as graphs where the node represents entities and the edges the connections between the entities. This kind of relationship mapping preserves the underlying properties of the data. Examples includes friendship network, telephone call network, co-authorship network, biological network, molecules, financial network and disease transmission. A special family of ML model called graph neural network (GNN) have been designed to handle such data.

    The problem though is that ML model tends to learn more than is required and as such, they leak more information when the model is released. This makes it a honeycomb for attackers to exploit. Such attacks includes membership, attribute and property inference attack.

    The drawback in all proposed attacks for ML model is that they are designed for Euclidean or independently and identically distributed (idd) data. However, GNNs utilize the aggregation of the features of the neighboring nodes to make prediction for the node. This makes it a unique problem as well as an opportunity to understand the risk posed by GNN models based on their utilization of the graph structure to make predictions. The aim of this master thesis is to investigate the vulnerabilities of different GNN models to different attacks.

    Prerequisites:

    1. Good knowledge of ML and Graphs
    2. Strong programming background in Python and libraries such as PyTorch

    References:

    • Iyiola E Olatunji, Wolfgang Nejdl, and Megha Khosla. “Membership inference attack on graph neural networks”. In:arXiv preprint arXiv:2101.06570 (2021)
    • Reza Shokri et al. “Membership Inference Attacks Against Machine Learning Models”. In: 2017 IEEE Symposium on Security and Privacy (SP). 2017, pp. 3–18.
    • Ahmed Salem et al. “Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models”. In: arXiv preprint arXiv:1806.01246(2018).
    • S. Yeom et al. “Privacy Risk in Machine Learning: Analyzing the Connection to Overfitting”. In:2018 IEEE 31st Computer Security Foundations Symposium (CSF). 2018, pp. 268–282.

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Megha Khosla Emmanuel Iyiola Olatunji

  • Identifying conserved host response for virus infections such as Covid-19 (Prof. Nejdl, Prof. Li)

    Clinical presentations of COVID-19 are highly variable, and while the majority of patients experiences mild to moderate symptoms, 10%–20% of patients develop pneumonia and severe disease. We recently performed the first single-cell RNA-sequencing of blood cells to determine changes in immune cell composition and activation in mild versus severe COVID-19 over time.1 A recent study based on multi-cohort analysis of host immune response identifies conserved protective and detrimental modules associated with severity across viruses2,3. Therefore, we hypothesized that viral infections induce a conserved host response and the conserved response is associated with disease severity. In this project, we will 1) implement the analysis framework for identifying conserved host response and 2) apply it to transcriptome data from multi-cohorts of COVID-19 and other virus infected patients.

    PDF-Flyer

    Leitung und Ansprechpartner der Abschlussarbeit

    <link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme professorinnenprofessoren wolfgang-nejdl u-icon-after>Prof. Dr. techn. Wolfgang Nejdl Prof. Dr. Yang Li

  • A benchmark for biomedical understanding (Tam Nguyen)

    Psychiatric Disorders (PDs) rank 5th in terms of prevalence and account for 6.7% of “Disability Adjusted Life Years”. We have explored different types of datasets to understand the landscape of research on psychiatric disorders. In particular, we designed a categorization of psychiatric data, including genomics data, molecules data, drug review data, research publication data, clinical data, etc. Moreover, we also built a repository of related venues and downable resources, including top-tier journals (Bioinformatics, ACM Transactions on Computing for Healthcare, Nature Research, Genome Research, Social Psyhiatry and Psychiatric Epidemiology, etc.), top-tier conferences (IEEE International Conference on Bioinformatics and Biomedicine, ACM Conference on Health, Inference, and Learning etc.), and top-tier workshops (AI for public health, BioKDD, WWW AI In Health). In this project, we will develop a benchmark based on the explored datasets and related to the aforementioned research community such as performing question answering to help users cross-check psychiatric facts.

    Possible benchmark topics include:

    1. A Benchmark for (bio)medical Machine Reading Comprehension
    2. Benchmarking the Generality and Domain-Specification of MCR models
    3. Benchmark on active learning for biomedical document annotation
    4. Benchmark on active learning for biomedical image segmentation
    5. Benchmark on active learning for biomedical question answering

     

    An ideal candidate should have:

    • Good background in machine learning, deep learning, and programming (Python or R).
    • Knowledge about data analytics, especially on social media, textual data, and knowledge graph data.
    • Experiences with natural language processing tasks such as machine reading comprehension and question answering is a plus.
    • Love to read and explore scientific articles, preferably every day.
    • Pro-active in learning new things, preferably every day.

    Interested students are encouraged to email to Dr. Tam Nguyen tamnguyen(at)l3s(dot)de for discussions.

    References:

    • Survey Paper 1: Q. Jin, Z. Yuan, G. Xiong, Q. Yu, C. Tan, M. Chen, S. Huang, X. Liu, S. Yu, Biomedical Question Answering: A Comprehensive Review, (2021).
    • Towards Medical Machine Reading Comprehension with Structural Knowledge and Plain Text (EMNLP 2020)
    • BioMRC: A Dataset for Biomedical Machine Reading Comprehension 

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Dr. Tam Nguyen

  • Debunking Medical Misinformation - Rootcauses, Benchmarks, and Explanations (Tam Nguyen)

    An abundance of false or misleading medical information has been observed on the Web and particularly on social media, posing a considerable threat to public health while eroding trust in healthcare systems. 6 out of 10 people search for the cause of their medical condition online, and among those who found a diagnosis online, 35% does not visit a professional medical provider. The COVID-19 pandemic has exacerbated this problem by bringing forward an infodemic surrounding the coronavirus that spreads as quickly and deadly as the virus itself. For example, rumours about remedies such as methanol to cure COVID-19 resulted in 300+ deaths and 1000+ people fallen ill. This project aims to build a framework to combat medical misinformation with focus on social media analytics, detection benchmarks, knowledge intensive tasks (question answering, machine reading comprehension, knowledge graph construction) and explainable mitigation measures.

    An ideal candidate should have:

    • Good background in data analytics, especially on social media, textual data, and knowledge graph data.
    • Motivated in learning and exploring the data of interest in some human level annotation.
    • Knowledge about machine learning, deep learning, and programming (Python or R).
    • Love to read and explore scientific articles, preferably every day.
    • Pro-active in learning new things, preferably every day.

    Interested students are encouraged to email to Dr. Tam Nguyen tamnguyen(at)l3s(dot)de for discussions.

    References:

    • Waszak, P.M., Kasprzycka-Waszak, W. and Kubanek, A., 2018. The spread of medical fake news in social media–the pilot quantitative study. Health policy and technology, 7(2), pp.115-118.
    • Naeem, S.B., Bhatti, R. and Khan, A., 2020. An exploration of how fake news is taking over social media and putting public health at risk. Health Information & Libraries Journal.
    • Treharne, T. and Papanikitas, A., 2020. Defining and detecting fake news in health and medicine reporting. Journal of the Royal Society of Medicine, 113(8), pp.302-305.

     

    Leitung und Ansprechpartner der Abschlussarbeit

    Tam Nguyen

  • Online Bahnanpassung für die intelligente Prozessregelung (Svenja Reimer)

    Bisher werden die Bearbeitungsbahnen bei der Fräsbearbeitung im NC-Code fest vorgegeben. Intelligente Überwachungssysteme können Prozessinstabilitäten bei schlecht gewählten Schnittparametern zwar bereits erkennen - eine Regelung der Schnitttiefe und -breite ist durch die fest vorgegebenen Werkzeugbahnen jedoch bislang nicht möglich. Der Inhalt dieser Arbeit ist daher die Entwicklung von Methoden zur autonomen Bahnanpassung.

    Teilaspekte der Arbeit sind:

    • Entwicklung von Algorithmen zur Bildung von parametrischem NC-Code
    • Erarbeitung von Grenzwerten und intelligenten Regelungsansätzen
    • Versuche in der Maschinensimulation und an der realen Werkzeugmaschine

    PDF-Seite

    Leitung und Ansprechpartner der Abschlussarbeit

    Svenja Reimer

  • Intelligente Prozessregelung durch mitlernende Stabilitätskarten (Svenja Reimer)

    Ratterschwingungen im Zerspanprozess führen zu einer schlechten Oberflächenqualität. Das Auftreten von Ratterschwingungen ist eng mit den gewählten Prozessstellgrößen (Schnitttiefe, - breite, Drehzahl) verknüpft. Zur Ermittlung der Zusammenhänge zwischen den Prozessstellgrößen und Prozessstabilität unter zur Wahl optimaler Prozessstellgrößen sind bislang aufwändige Simulationen notwendig. Durch den Einsatz von maschinellem Lernen und modernen Überwachungssystemen können diese Zusammenhänge auch im Prozess selbständig von der Maschine erlernt werden ("mitlernende Stabilitätskarten"). Ziel dieser Arbeit ist die Entwicklung und Umsetzung einer intelligenten Prozessregelung auf Basis der mitlernenden Stabilitätskarten. Die Arbeit umfasst unter anderem folgende Punkte:

    • Entwicklung von Bearbeitungsstrategien zur gezielten Datengenerierung für mitlernende Stabilitätskarten
    • Entwicklung von Strategien für das "online-Lernen" während dem Prozess
    • Entwicklung einer Zielfunktion für die Regelung
    • Zerspanuntersuchungen zur online Anpassung von Prozessstellgrößen

    PDF-Seite

    Leitung und Ansprechpartner der Abschlussarbeit

    Svenja Reimer

  • Understanding Relevance Search over Medical Knowledge Graphs (Prof. Ganguly)

    There is an exponential rise in the amount of medical evidence being produced, which makes it very difficult for medical professionals to stay regularly updated with the recent research studies in order to practice evidence-based medicine. In this master thesis, we aim to accommodate richer query variations for online biomedical literature search and redesign the document collection used for search into a knowledge graph. In other words, we will adapt the “exemplar query” setting developed by Mottin et al. (2016,2018) to the biomedical information retrieval domain and try to achieve state-of-the-art performance in TREC Precision Medicine Track.

    Specifically, the student will get first-hand experience working with unstructured text (clinical notes) and knowledge graphs, in a medical domain. 

    Prerequisites:

    1. Good knowledge of Natural Language Processing (NLP) and Graphs
    2. Strong programming background in Python and working knowledge of standard Machine Learning and NLP libraries

    References:

    1. Mottin et al. Exemplar queries: a new way of searching (VLDB 2016)
    2. Gu et al. Relevance Search over Schema-Rich Knowledge Graphs (WSDM 2019)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Niloy Ganguly Soumyadeep Roy

  • Fairness-aware Online Learning under Class Imbalance (Dr. Vasileios Iosifidis, Prof. Dr. Wolfgang Nejdl)

    Fairness-aware online learning has become an evolving field during the last fewyears. Fairness-aware online learning goal is to maintain a classifier that performswell and does not discriminate over the course of the stream. Some initial works havebeen proposed to tackle discriminatory outcomes from online classification [1, 2];however, these methods do not take into consideration the uneven class distributionover the course of the stream. If the imbalance problem is not tackled, the learnermainly learns the majority class and strongly misclassifies/rejects the minority. Suchmethods might appear to be fair for certain fairness definitions that rely on parity inthe predictions between the protected and non-protected groups. In reality though thelow discrimination scores are just an artifact of the low prediction rates for theminority class.

    In this master thesis, we want to investigate the combined problem of class-imbalanceand fairness-aware learning in the online setup. We focus on Naive Bayes classifierwhich has been extensively studied in the context of fairness but in the static setting.In this work, we plan to extend these models to the online setting taking into accountthe imbalance of the population under different fairness notions such as statisticalparity[3], equal opportunity [4], and equalized odds [4].

    An ideal candidate should be:

    • a self motivated and independent learner
    • knowledgeable about machine learning (good grades in Data Mining I, DataMining II)
    • experienced with python or java

     

    Interested students are encouraged to email to Wolfgang Nejdl and/or Vasileios Iosifidis at for scheduling an appointment. CV and transcript of records must be sent beforehand.

    References

    1. V. Iosifidis, H. Tran, E. Ntoutsi, "Fairness-enhancing interventions in streamclassification", 30th International Conference on Databases and ExpertSystems Applications (DEXA), 2019.
    2. W. Zhang, E. Ntoutsi, "An Adaptive Fairness-aware Decision Tree Classifier",International Joint Conference on Artificial Intelligence (IJCAI), 2019.
    3. Kamiran, F., & Calders, T. (2012). Data preprocessing techniques forclassification without discrimination. Knowledge and Information Systems,33(1), 1-33.
    4. Hardt, M., Price, E. and Srebro, N., 2016. Equality of opportunity insupervised learning. In Advances in neural information processing systems (pp.3315-3323).

     

     

    Leitung und Ansprechpartner der Abschlussarbeit

    <link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme wiss-mitarbeiterinnenmitarbeiter vasileios-iosifidis u-icon-after>M.Sc. Vasileios Iosifidis<link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme professorinnenprofessoren wolfgang-nejdl u-icon-after>Prof. Dr. techn. Wolfgang Nejdl

  • Question Answering Using Deep Learning (Prof. Anand)

    Leitung und Ansprechpartner der Abschlussarbeit

    <link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme junior-professorinnen-professoren avishek-anand u-icon-after>Prof. Dr. Avishek Anand

  • Faster Inference for Deep Neural Rankers (Prof. Anand)

    Leitung und Ansprechpartner der Abschlussarbeit

    <link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme junior-professorinnen-professoren avishek-anand u-icon-after>Prof. Dr. Avishek Anand

  • Interpretability of Neural Models (Prof. Anand)

    Leitung und Ansprechpartner der Abschlussarbeit

    <link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme junior-professorinnen-professoren avishek-anand u-icon-after>Prof. Dr. Avishek Anand

  • Neural Information Retrieval (Prof. Anand)

    Leitung und Ansprechpartner der Abschlussarbeit

    <link de kbs team personen-detailansicht epv fachgebiet-wissensbasierte-systeme junior-professorinnen-professoren avishek-anand u-icon-after>Prof. Dr. Avishek Anand

  • Modelling and predicting the knowledge state of students (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Search as learning (SaL): Investigating the impact of images and videos in SaL scenarios (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Automatic highlighting of importants segments in educational videos (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Automatic question generation for (educational) videos (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Using knowledge graphs for video question answering (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Automatic captioning for scholarly figures (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Analysing and Linking Graphical Representations in Computer Science Publications to their Implementation (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Linking Formulas in Scientific Publication with their Software Implementation (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Information extraction from scientific (textual) publications for knowledge graph enrichment (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Improving OCR for recognizing text in scientific videos (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth

  • Semi-automatic enrichment of scientific videos with external recommendations (Prof. Ewerth)

    Leitung und Ansprechpartner der Abschlussarbeit

    Prof. Dr. Ralph Ewerth