Available Theses-Internships

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis or Gerhard Weiss):

Machine learning (especially deep learning techniques) have recently found their way to mimic existing artistic styles or judging pictures, paintings for their aesthetic quality. While the inner details of the algorithms are not directly interpretable, a question that arises is whether the algorithms learn something about concepts like “harmony”.

In this thesis you will work with a real artist (that will provide us with enough data) and explore (a) if there is a way to build such a system that will identify whether an artistic synthesis is in harmony or not and (b) whether after having “seen” many examples the system is able to generate new syntheses that are in harmony. We are particularly interested in assessing the characteristics that lead the algorithm to make specific decisions. Symbolic learning (feature extraction, etc.) and deep learning (feature discovery, etc.) techniques will be explored.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis ):

Dialogue systems are emerging as a hot application area of deep learning architectures. There are many examples of how deep learning can be applied to either general domain question-answering or to specific domains (e.g. the ubuntu corpus, company customer support). The main goal is that neural network has to model questions & answers and also keep track of the dialogue state. There is already a developed system (based on the VHRED architecture) that for now provides answers based on information retrieval techniques (i.e. the answer that best matches the question).

Your goal in this thesis would be to extend current model by fine-tuning its architecture and give the possibility to the "chatbot" to answer specific questions. There is also the opportunity of working with a specific corpus of customer support dialogues.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis ):

Convolutional neural networks have been extensively and successfully studied and applied to image data. Convolution is a powerful technique from signal processing that in the concept of images is applied by siding a small kernel matrix over all possible positions in the image with the goal of identifying important features. However, convolution (“optical flow”) can be applied to other data (e.g. temporal) usually in 1-dimension and can show how different variables convolve over time.

Given a multivariate (numeric and categorical) dataset spread over discrete time steps, your goal in this thesis would to be to study how 1-dimensional-convolution can be applied to a prediction task (research into how variables can be represented and how convolution windows can be applied) and construct a NN architecture that solves the problem.

Proposed Master (AI or OR) Thesis (contact Dr Stelios Asteriadis):

Transfer learning is an important open issue in the field of machine learning. The exchange of knowledge allows for a better and faster training of new systems using “experiences” acquired through other ones.

We are looking for a student with good programming skills (preferably Python and/or Java) interested in developing their master thesis in this promising and highly rising area in AI. The main objective of the project is to develop a general AI to be applied in data acquired through different sensors during interacting with serious games. The method to be implemented must allow for knowledge sharing between different games, in order to improve the overall experience provided by the interaction. Transfer learning will facilitate the training of a device which uses, for instance, video data for emotion elicitation, using the information acquired using a different device through, e.g., audio or inertial sensor data (how computer vision based emotion recognition can benefit emotion recognition from other sensors).

The outcome of this project will be the implementation of a method to modify game parameters (e.g. difficulty level) to adapt the game to user performance (e.g. emotional state). This thesis will be focused on transfer learning and it won’t start from scratch. Initially, some datasets and analysis algorithms will be provided in order to facilitate the implementation of the system.

Key words: transfer learning, machine learning

Level: Master in AI or OR. Good programming skills (preferably Python and/or Java) and knowledge about machine learning

Recommended literature:

Zhu, Y. Chen, Z. Lu, S. Jialin P., G. R. Xue, Y. Yu, Q. Yang, Heterogeneous Transfer Learning for Image Classification, AAAI, 2011
T. Zhou, S. J. Pan, I. W. Tsang, Y. Yan, Hybrid Heterogeneous Transfer Learning through Deep Learning, Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence

Proposed Bachelor or Master (AI or OR) Thesis (contact Dr Enrique Hortal):

Mobile phones are probably the first device you use every morning and the last one you use before going to bed. It is becoming a “mandatory” gadget, as it can provide us with a huge amount of information from different sensors. The objectives of this project are the following:

Integration of a serious learning game, already developed by our research group, in mobile (android and/or iOS) devices making use of touch screen-based interaction
Affect analysis using sensors embedded in the mobile device (e.g. gyroscopes and touch screen data analysis). Ready-to-use tools and algorithms, available in our research group, will be provided for data analysis.

Key words: mobile sensors, affect analysis, serious game

Level: Bachelor or Master (AI or OR). Good programming skills (preferably Python and/or Java) and knowledge about machine learning

Recommended literature:

Céline Coutrix and Nadine Mandran. "Identifying emotions expressed by mobile users through 2D surface and 3D motion gestures." In Proceedings of the 2012 ACM Conference on Ubiquitous Computing, pp. 311-320. ACM, 2012.

Proposed Bachelor or Master (AI or OR) Thesis (contact Dr Stelios Asteriadis):

Exchanging experiences with our peers during learning is a very important pedagogical technique. The collaboration between learners with different learning profiles and records implies a new student-student relationship (in contrast to the traditional, student-teacher one) where both students benefit from each other (asking one another for information, evaluating one another's ideas, monitoring one another's work, etc.).

This thesis aims at creating a collaborative game where several students can work together using different technologies (i.e. computers and mobile devices). To that end, the student will adapt a serious game, already developed in our group, to be easily used in mobile (Android and/or iOS devices) devices, while, at the same time, sharing goals and performance with peers, in real time. User Experience tests are also expected, following Human-Computer Interaction principles.

Key words: serious games, collaborative games, sharing information

Level: Bachelor or Master (AI or OR). Good programming skills (preferably Python and/or Java) and elementary knowledge about machine learning

Proposed Bachelor or Master (AI or OR) Thesis (contact Dr Enrique Hortal):

Data annotation is a tedious process that usually implies the expenditure of considerable time, effort and resources. Using already recorded data, this thesis will implement a tool to extract key moments from face videos using machine intelligence. These videos come from experiments of people interacting with a serious game and key moments are expected to correspond to key events in the interaction.

The main objective of this project is to implement a simple tool to obtain such video frames which contain useful information about the affective state of the user. To do that, the student will use machine learning in order to differentiate between representative key moments and (more) neutral phases of the game (e.g. using clustering techniques).

To validate this data, manual annotation tools will also be implemented, allowing for humans to extract key moments in people’s expressivity. This project requires good programming skills (preferable Python and/or Java) and some knowledge about machine learning.

Key words: annotation tool, automatic video annotation, machine learning, serious game

Level: Bachelor or Master (AI or OR). Good programming skills (preferably Python and/or Java) and elementary knowledge about machine learning

Recommended literature:

J. Rich and M. Hannafin, (2009). Video annotation tools: Technologies to scaffold, structure, and transform teacher reflection. Journal of Teacher Education, 60, 52-67.

Y. Lin, B. Tseng, and J. R. Smith, “VideoAnnEx: IBM MPEG-7 annotation tool for multimedia indexing and concept learning,” in Proc. Int. Conf. Multimedia & Expo, Baltimore, MD, 2003.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

The increasing amounts of text documents (like news items or articles) makes it impossible to follow the trends, events or narratives or the developing storylines. Given a topic detection algorithm applied on a large news corpora dataset (Reuters archive from 1996 and 2015 are available in RAI), the goal of this project is to develop NLP/information retrieval techniques that will facilitate the storyline presentation given the underlying hidden topic structure. Emphasis will be given to the evaluation measures of the results.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

Word embeddings is the latest and most effective way to represent documents but has only been extensively applied to English language texts. Recently, embeddings for dutch language were presented (trained on a large corpus). Goal of this thesis is to utilise this representation along with deep learning techniques (like Convolutional Neural N etworks) in order to classify a corpus of (dutch) short texts to some predefined categories.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

Graduating attributes (or competences) describe the sets of values that students should develop by the end of their studies. Based on the courses content and students’ development, in this project you will explore different techniques (tensor decomposition, matrix factorization, word2vec, etc.) in order to build a course recommender system that will suggest students’ courses that will improve either their average competence profile or a specific competence.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

Graduate student profiles can be described by many attributes: grades, courses curriculum or even hidden dimensions (like their competences). Similarly, job advertisements can be represented by sets of values that should match those of students’ profiles. Goal of this thesis is to build a reciprocal recommender system that will match students to jobs (and vice versa). Evaluation measures of the algorithm and the overall proposed matching will also be explored.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

Online learning through Massive Open Online Courses (MOOCs) is an emerging education technology area with increasing demands. However, it is found that out of the many thousands of participants enrolled in various MOOC courses, the completion rate for most courses is below 13%. There is not much knowledge yet on the profiling of people taking MOOCs, in order to predict dropout rates. Goal of this thesis is to study a MOOC dataset (already available) and identify or predict dropout probabilities based on user behavior characteristics.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

On-line social networks have become a massive communication and information channel for users world-wide. Twitter is a means used by many people in order to express their opinions, complaints, thoughts, etc. for almost every topic. Despite many efforts, there is no clear or robust way to evaluate this kind of data yet (e.g. if a specific tweet corresponds to reality or not or if it corresponds to an accurate review).

Goal of this thesis is to pick a topic (of your interest, otherwise there is already an idea), collect relevant to this topic tweets (based on Twitter streaming API) and then explore or develop methods from literature that can be used to extract useful information. Emphasis will be given to the evaluation of these techniques.

Proposed Bachelor/Master Thesis (contact Gerasimos (Jerry) Spanakis):

Airport congestion and scarce capacity at popular airports are rising problems with no straight-forward solutions till now. An airport slot is a right granted by an airport owner which allows an airline to schedule a landing or departure during a specific time period. Allocated slots may have a commercial value and can be traded between airlines.

Goal of this thesis is to expand an already implemented auction-based system for airport slots in order to further improve the allocation of slots according to specific criteria (profit maximization of airlines, airport utilization, etc.). Dynamic pricing techniques can also be explored as well as airline behavior/negotiation models.

Proposed Master Internship / Thesis (contact Gerhard Weiss):

In close cooperation with and co-supervised by:
Marco Das, MD, PhD, MBA, Head of CT Department
Department of Radiology, Maastricht University Medical Centre (MUMC)

The improvement of workflows is crucial for the efficiency and competitiveness of companies and any other kind of organizational units. This internship/thesis is about workflow improvement in hospitals. Specifically, and in cooperation with MUMC, it is about improving the current workflow from clinical request to the final report of a patient. This workflow is complex and requires multiple tasks to be fulfilled with many manual tasks from different employees (administration, CT technician, radiologist) as well as multiple different computer systems (Hospital Information System SAP, Administration tasks Qdoc, CT scanner software, reporting in AGFA Pacs). In its current form this process is inefficient and even may lead to unsatisfying results for the patient (increased waiting time, wrong protocol, overlooked pathology due to insufficient imaging).

The task is to develop a software for an efficient workflow, which combines necessary tasks in terms of administration, as well as optimal patient routing and planning and decision making at which scanner with which protocol the patient has to be scanned. The ultimate goal is to have the final report of a patient with the optimal imaging in the shortest amount of time.

Proposed Bachelor Thesis / Internship (contact Kirill Tumanov):

In this thesis project you will be exploring how to design, build and evaluate a Hidden Markov Model (HMM), applied to brain signal analysis. Markovian approach assumes that there is a certain state regularity observed in the analized data. But what exactly is this pattern? So far HMMs were extremely successful for instance in speech and handwriting recognition. It has also been found useful in bioinformatics and brain signal temporal analysis.

At the moment a standard Support Vector Machine (SVM) classification is used at UM when working with functional Near-InfraRed Spectroscopy (fNIRS) system for brain recording. SVM will be a banchmark for you to test the designed HMM performance against. Recent research indicated a high potential of HMM to replace SVM in this domain. You will be able to test if this is the case, while deepening your knowledge in Markovian approaches in general and the HMM in particular.

Proposed Master Internship or Thesis in RAI, in collaboration with Qwiek (http://qwiek.eu/) (contact Stylianos (Stelios) Asteriadis, Gerhard Weiss, Rico Möckel ):

The Qwiek.play is a gaming computer for elderly in care homes. The Qwiek.play uses a combination of television, tablet (iPad), RGB camera and a user centered interface to stimulated elderly in a playful, attractive but mainly active way. Games are especially designed to fit their interest and are focuses on their personal perception. Currently we are using the RGB camera to detect motion and face detection (using OpenCV), this works well for the current games. But, we believe that we can get a lot more out of this RGB camera without switching to an expensive 3D camera (and this is why we need you). For instance, if we make the camera more intelligent, we would be able to detect patterns. This can be the detection of simple events like; counting how many people stand in front of the system, or how many people have passed by today or slightly more advanced: detect gestures like waving, raised hands or even a throwing gesture. Your task will be to analyze the camera feed and detect these events (and preferably come up with a lot more).

Step two of the project is more about an user adaptive interface. What do we mean by this? The patterns, which we have stated above, can be used as input for the system to respond “intelligent” in certain situations. But to be truly “intelligent” the system needs to determine the kind of “mood” a user is in. Is he/she laughing or actually being annoyed or bored. Being able to determine these events through image processing, application usage data or sound and let the application respond to this accordingly, would be a very valuable addition to the product. The end goal would be a system which is able to measure its own user interest and make sure it keeps itself interesting for a specific user.

Proposed Master Internship or Thesis in RAI, in collaboration with Qwiek (http://qwiek.eu/) (contact Stylianos (Stelios) Asteriadis, Gerhard Weiss, Rico Möckel ):

All smartphones which are being sold today are equipped with a 9DOF sensor. A 9DOF is (mostly) a 3-axis accelerometer combined with a 3-axis gyroscope and a 3-axis magnetometer (compass). The combination of these sensors are, in theory, able to detect accurate 3D movement of an object in space. Knowing the 3D movement/orientation of an object could be interesting for a variety of applications (for example controlling a drone), but we the area that we are very interested in is measuring accurate body movement (when the sensor is attached to a limb).

Why? Therapist are very interested in analysing the development of a revalidation process. They can measure the progress (or deterioration) and check whether clients are performing a task correctly. Furthermore we can also use the data from the sensor to control training games (serious gaming).

So what are we looking for? To achieve this accurate measurement we need to develop the correct algorithm and noise reduction. You won’t start from scratch! There are a lot of sample projects online and we have made some fairly good progress ourselves. A successful project outcome would be a system which is able to detect reliable body movement within a 3D space and can learn to detect a specific exercise coming from a wide variety of people (with different body dimensions) and give insights in the quality of the movement.