2023 PhD Recruitment

Join the 2023 cohort of Pioneer Centre for AI PhD students – we expect 15+ new students to start in the fall, across the centre’s partner universities. Each student will be employed at one of the partner universities AND become a part of the Pioneer Centre for AI. PhD students at the Pioneer Centre for AI have extraordinary access to computing resources, international researchers across many disciplines within computer sciences and other academic areas, as well as courses and events at the centre, and meaningful collaboration with industry, the public sector, and the start-up ecosystem.

To apply for one or more of the positions, follow the link after the individual project description. Note that the application deadlines may vary slightly, April 1st or 2nd. The Aarhus University deadline is May 1st. 

You can read more about the Master’s and PhD study in Denmark here>>. 

• XR: Human–Computer Interaction with Virtual Reality, University of Copenhagen (deadline: 1 April 2023): The PhD fellow will work on human–computer interaction with virtual reality. The position is part of a project that investigates movement learning in VR. The project aims to theorise how sensory feedback (e.g., visual or haptic) influences moving in VR (e.g., redirection), and to optimise motor learning through designing how user interfaces for VR provide feedback. Therefore, knowledge and interest in human–computer interaction research is essential, but as a PhD student you have much freedom to detail your focus: to a large extent, you shape the research you work on and the ideas you pursue. We seek a candidate who is passionate about human-computer interaction research and experienced in one or more of the following: empirical studies of human–computer interaction, motor learning, or perception; or designing or developing software or hardware prototypes of interaction techniques or user interfaces. You should demonstrate a keen interest in virtual reality, movement sciences, cognitive sciences, multimodal interaction, or other topics related to the project.
  Supervisor: Joanna Bergström.
  Find more information and link to application here>>.

• Collaborative Interfaces for Mixed Reality, Aarhus University (deadline: 1 May 2023). How will we collaborate in the future? Virtual (VR), Augmented (AR) and Extended Reality (XR) is increasing in popularity as a new paradigm for remote collaboration. In the Metaverse, future collaborative applications enable new forms of creative meetings where people can break out of their boring Zoom windows and be together in a shared 3D world.
  This project will conduct foundational Human-Computer Interaction (HCI) research to advance the state-of-the-art of XR collaborative interactions. This involves, but is not restricted to, the design, implementation, and evaluation of novel interaction concepts and systems, to better understand and shape the future of XR collaboration. The project can include, but is not limited to: 
  • Mixed physical-real collaboration: Multiple persons can be set in different physical locations/rooms/places, be co-located or remote, which raises the question how to integrate physicality in the collaborative scene.
  • AI-driven collaboration: XR technology provides a wealth of contextual information that can be used to predict user action, and adapt the UI to the users.
  • Hybrid work and meetings in XR: People can enter collaborations from various devices (mobiles, head-worn, etc.), and it is an open question how to best support users with different devices.
  • Multimodal collaboration: Multimodal XR with e.g. hand, face, and eye tracking offer realistic collaboration opportunities close to real-world meetings, but also provide novel ways to design collaborative interfaces that harness the flexibility of virtually-mediated communication.

  Supervisor: Ken Pfeufer and co-supervisor: Hans Gellersen.
  Find more information and link to the application here >>.

• White matter lesion and acute stroke detection in the wild, University of Copenhagen (deadline: 1 April 2023): We are looking for a candidate with keen interest in Machine Learning, Deep Learning, Pattern Recognition, and/or Medical Image Analysis. This project aims to identify and characterize the brain’s white matter abnormality and acute stroke in MRI scans by developing robust deep/machine learning methods with fine-grained conclusions drawn from standard clinical data harvested in the wild to help with the disease diagnosis. The robustness will be proven to variations in scanner vendor, field strength, software, quality, and protocol, on a very heterogeneous data with natural variation of comorbidities and demographics.
  Supervisor: Mads Nielsen. Co-supervisor: Mostafa Mehdipour Ghazi.
  Find more information and link to application here>>. 

• Externalist XAI, University of Copenhagen (deadline: 1 April 2023): Externalism is the mainstream position in natural language semantics, but large language models and related technologies are commonly analyzed using internalist techniques.This PhD will aim to extend current XAI taxonomies to include externalist approaches, discuss how they relate to functionalist accounts, and – through a combination of theoretical and empirical analysis – investigate the pros and cons of externalist accounts. Research questions include: Are functional approaches to XAI generalizations of externalist approaches, and what criteria would be further satisfied by externalist accounts? Is there a continuum from functional approaches to internalist approaches? Can externalist accounts provide knowledge or understanding that internalism cannot account for? Finally, does the analysis of AI systems shed new light on human communication, e.g., by deanthropomorphizing current accounts?
  Supervisor: Anders Søgaard. Co-supervisors: Herman Cappelen (University of Hong Kong), and Thor Grünbaum.
  Find more information and link to application here>>.

• Uncovering the fundamental limits of ML and AI for predicting human behavior,  IT University of Copenhagen (deadline: 1 April 2023): Is everything predictable given enough data and computational power? There have been plenty of optimistic claims by journalists, researchers, and companies about the possibility to predict different phenomena from epidemics, the stock market, crime and even wars. However, research has found these claims to be exaggerated, demonstrating that social systems are notoriously difficult to predict. Nonetheless claims about predictability are generally believed by policy makes and the public. With the rise of algorithmic decision making and with automated systems mediating an increasingly larger part of our social, cultural, economical, and political interactions, it is vital to understand the limits of prediction and when predictive accuracies fall short of expectations. Are prediction limits determined by the size and bias present in datasets, the scale of computational power, or are there fundamental limits to prediction. To tackle this question the project will focus on three subgoals. 1) We will build on top of existing statistical frameworks to predict events driven by collective phenomena in social systems. 2) Develop an understanding of how biases in data affect predictive accuracies. 3) Use information theoretical methodologies to quantify the upper limit of predictability. This project focuses on prediction in social systems and aims to develop a more critical understanding what predictive systems can be used for, and where we need to be careful.
  Supervisor: Vedran Sekara and co-supervisor Roberta Sinatra.
  Find more information and link to application here>>.

• Modelling compositional language structures across languages, IT University of Copenhagen (deadline: 1 April 2023): The origin of language development of human beings can at least be traced back to two sources: the survival needs for communication and the unique wiring of human brains. The two factors can co-develop with each other through life-span. Nevertheless, languages from the whole world are with tremendous diversity. Yet, given the fact that all human brains share very similar structures and connection patterns, which presumably afford for similar ways to encode and decode information from the external world, there likely exist some hidden common structures, like some word orders or numbers of phonemes in a language, that are shared at least within, or even across, language families. An important constraint for these similar structures comes from the theory that the brain is processing language embodied in all our senses and via processing streams that are also involved in a range of other cognitive functions from specific motor control up to general problem-solving. This suggests that language comprehension and production, in fact, developed on top of existing information processing schemes, which in turn might have similarly shaped how the different language families have developed. A particular mechanism that was recently hypothesized to give rise to the structure of the brain’s sequence processing is temporal compositionality and chunking, which seemingly operate on language sequences as well. With this PhD project, we want to identify and describe the specific, latent temporal encoding structures that may constrain the temporal features of spoken language. In this project, the candidate will study structure patterns in spoken language and investigate how to build a model that can extract temporal characteristics of speech across different languages. The successful candidate will bring a background in either computer science, computational linguistics, or computational neuroscience, as well as an interest in the respective other disciplines.
  Supervisor Stefan Heinrich and co-supervisor:  Barbara Plank.
  Find more information and link to application here>>.

• Automatic Analysis of Mental Health by Machine Learning, IT University of Copenhagen (deadline: 1 April 2023): Mental disorders often go undiagnosed or misdiagnosed for years, including bipolar disorder, depression, autism, etc., negative impacts on individuals and communities. This PhD project aims to facilitate better understanding of mental disorders, by investigating the audio and video data of patient vs. control group, using tailored machine learning methods and developing them further. The long sighted goal is to potentially open the way to automatic diagnosis of mental disorders based on the new findings. Previous studies have shown that facial expressions contain some information about the mental health of a person, specifically that the emotional response differs between people with bipolar disorder and the healthy control group, which offers great prospects for further investigation. Previous approaches have employed neural networks, e.g. LSTMs (Long-Short-Term-Memory), coupled with Multi-Layer Perceptron networks on videos to classify mental disorders. While the results are encouraging, they mostly focus on the classification or prediction task without investigating any further. This project aims to go further by analysing different modalities of data, such as speech, eye movements, behaviour, virtual reality (VR) interactions. New Machine learning algorithms, tailored to one modality or combinations, should uncover new insights into the nature of mental disorders.
  Supervisor: Sami Brandt and co-supervisor Stella Graßhof.
  Find more information and link to application here>>.

• Improving Efficiency of NLP Models by Exploiting NLP Datasets, IT University of Copenhagen (deadline: 1 April 2023): Large language models have revolutionized the field of Natural Language Processing (NLP) since their introduction in 2018, leading to superior performance on almost all NLP tasks. These models are first pre-trained on enormous amounts of raw, unlabelled texts, and then fine-tuned for the target task of interest. However, due to their computational costs, training these models can currently only be done by large tech companies. Recently, it has been shown that language models can be improved by an intermediate step of training on a diverse set of NLP tasks. This second step of training is computationally cheaper compared to the initial pre-training paradigm, with normally around 1,000 times smaller data training data (millions of words versus billions). In this project we aim to investigate whether competitive performance can be obtained without the pre-training step. When training a language model only on a set of diverse NLP tasks, training costs will be only a fraction of current models.
  Supervisor: Rob van der Goot.
  Find more information and link to application here>>.

• The Business Value of Big Data Analytics and Machine Learning Algorithms, IT University of Copenhagen (deadline: 1 April 2023): By applying large-scale simulation and data analysis of live business datasets, this PhD project investigates how much data and which machine learning algorithms are most effective and sustainable in creating value across different business decision-making contexts. In doing so, the project will generate insights to inform decision-makers about the relevant scope of data and machine learning algorithms for specific business applications, thus making big data analytics more cost efficient, accessible, and sustainable. Research questions to be addressed include: How big do data need to be in order to create value in business decision-making? What is the relevant complexity of machine learning algorithms applied to business decision-making? What is the most sustainable application of machine learning algorithms and big data analytics to different types of decision-making contexts?
  Supervisor: Jonas Valbjørn Andersen.
  Find more information and link to application here>>.

• Bias Explained: Pushing Algorithmic Fairness in Science with Models and Experiments, University of Copenhagen (deadline: 1 April 2023): Data about popularity and impact are at the core of most measures and algorithms in society. Products are recommended based on user ratings. Trending news are featured on social media. In science, citation numbers are widely used to assess scientists and form the basis of recommendation engines like Google Scholar. However, such impact metrics reflect only a community’s reaction to a performance; they are not a measure of the performance itself. Because of their social nature, these metrics have one crucial issue: they build on biased data produced by people. For example, given equal performance, black/female researchers are less cited than white/male ones due to discriminatory citation cultures. Because many crucial decisions use the output of metrics and rankings, we ask: Given equal performance, what are the mathematical laws governing the evolution of inequality? What is the effect of the impact-based biases on the output of metrics and algorithms? To tackle these questions, this project plans to break new grounds by combining mechanistic models with controlled experiments, applied to large-scale datasets. This will push new frontiers in algorithmic fairness. 
  Supervisor: Roberta Sinatra.
  Find more information and link to application here>>.

• Self-Supervised Learning for Decoding of Complex Signals,  Aalborg University (deadline: 2 April 2023): To date, most successful applications of deep learning in signals and decoding are based on supervised learning. However, supervised learning is contingent on the availability of labelled data, i.e., each sample has a semantic annotation. The need for labelled data is a serious limitation to applications at scale and complicates the maintenance of real-life supervised learning systems. The typical situation is that unlabelled data is abundant, and this has given rise to paradigms such as semi-supervised and self-supervised learning (SSL). Both directions in SSL are based on combining large amounts of unlabelled data with limited labelled data. While semi-supervised learning invokes generative models to learn representations that support learning with few labels, self-supervised learning is based on supervised learning with a supervisory signal derived from the data itself. The goal of this PhD study is to develop novel semi-supervised and self-supervised methods for modeling signals of various modalities (e.g., speech, audio, vision, text) and analyse the complexity of the developed models.
  Supervisor: Zheng-Hua Tan and co-supervisor Lars Kai Hansen. 
  Find more information and link to application here>>.

• Self-Supervised Learning for Decoding of Complex Signals,  Technical University of Denmark (deadline: 2 April 2023): Are you passionate about machine learning and neuroimaging? To date, most successful applications of deep learning in signals and decoding are based on supervised learning. However, supervised learning is contingent on the availability of labeled data, i.e., each sample has a semantic annotation. The need for labeled data is a serious limitation to bio-medical applications at scale and complicates the maintenance of real-life neuroimaging systems.  Here we will combine international open source EEG (electric brain wave) data with local, high quality labeled data. The typical situation is that unlabeled data is abundant, and this has given rise to paradigms such as semi-supervised and self-supervised learning (SSL). Both directions in SSL are based on combining large amounts of unlabeled data with limited labeled data. While semi-supervised learning invokes generative models to learn representations that support learning with few labels, SSL is based on supervised learning with a supervisory signal derived from the data itself.
  Supervisor: Lars Kai Hansen and co-supervisor Zheng-Hua Tan.
  Find more information and link to application here>>.