Research projects- AI4Research


Eight research projects associated with artificial intelligence and machine learning have been awarded funding within AI4Research. Read more about the projects below.

The Potential and Limits of AI

David Sumpter, The Department of Information Technology

There is increasing evidence that AI is being overhyped in societal application areas, in particular, but also in medical areas. Although the question of whether AI is over-hyped is simple to state, there is no single straightforward answer. Instead, what is required is for us to build capacity, both in industry and academia, for using AI in a way which utilizes its genuine benefits, but doesn’t exaggerate what it can do.

This research proposal will take stock of this increasing evidence coming from application areas, in the context of the meta-lessons I have learnt from my own research, teaching and industrial collaboration the last years. It will also look at how the hype is affecting AI Research in Uppsala. The project will:

  • Start to categorise best AI practice and document how some practices can lead to misleading claims.
  • Look at now progress on benchmark problems gives incorrect impressions of advances.
  • Investigate ways in which data-centric machine learning can advance.
  • Investigate ways of measuring performance and negative impact of AI.

I believe that it is important that researchers, like me, who have experience with the techniques that now constitute AI take a pause and reflect over the next steps for our research area.

What do you look forward to the most during your sabbatical?

Meeting and talking to other researchers about how they are using AI. Part of my project is to conduct structured interviews with others work at AI4Research, to find out what potentials and limits they see in their work. So much happened in the field during the pandemic and it will be nice to have those daily contacts with people working in this area, to see how they experienced the changes.

Machine Learning Quantum Gravity

Magdalena Larfors, Department of Physics and Astronomy

As a theory of quantum gravity, string theory may be used to solve outstanding questions in physics ranging from microscopic to macroscopic scales. In recent years, machine learning (ML) has emerged as a useful tool in this research. For example, ML tools has allowed us to efficiently construct semi-realistic string models for physics, thus connecting string theory to the real world. ML techniques appear particularly apt to explore the geometry underlying string theory, and has the potential to reveal new mathematical results. Conversely, string theory offer a great setting for experiments in ML: examples are abundant, large data sets can be generated at ease, and constraints from physics, mathematics and geometry guide explorations. The aim of this project is to further promote this interdisciplinary research. It will consolidate and expand new open source ML packages for the construction and analysis of (semi-)realistic solutions of string theory, which may be explored to derive new physics-inspired ML results.

What do you look forward to the most during your sabbatical?

To work in an interdisciplinary environment, expand my knowledge in machine learning and data science, use this knowledge in my research, and establish new interdisciplinary collaborations.

Magdalena Larfors

Magdalena Larfors

Senior Lecturer/Associate Professor at Department of Physics and Astronomy, profile page

A Brain-Inspired Learning Framework for Data and Energy Efficient AI

Ayca Ozcelikkale, The Department of Electrical Engineering

Success of artificial intelligence (AI) methods typically require large computational times and energy for computation. Recently, computing on neuromorphic chips has shown gains up to 100000× in terms of energy-delay product for certain applications compared to the traditional chips. Despite these promising initial results, extending these successes to a wider range of applications is a major challenge due to limitations of learning algorithms applicable on neuromorphic computing platforms. This project focuses on this challenge. Our main tool is spiking neural networks (SNNs). SNNs process data using spikes similar to how our brains process information. Using SNNs, we will develop machine learning solutions compatible with neuromorphic hardware. The goal of the project is to contribute to data and energy efficient AI using these brain-inspired solutions.

What do you look forward to the most during your sabbatical?

I look forward to exchanging ideas and discussing the future of AI with other researchers.

Deep Learning for High-Energy Neutrino Astronomy

Christian Glaser, Division of high-energy physics, department of physics and astronomy

Deep learning has the potential to boost the emergent field of high-energy neutrino astronomy substantially. I plan to use deep learning in multiple areas, ranging from low-level trigger optimizations (where execution time and power consumption is the limiting factor, like typical IoT and real-time-analysis challenges) over event reconstruction with large deep neural networks towards an end-to-end detector optimization using differential programming. This can enhance the science capabilities of future neutrino detectors substantially by increasing the detection rate of neutrinos and improving the quality of each detected event. The timing of this project is perfect for influencing IceCube-Gen2 - the largest facility for astroparticle physics for the next decade - whose construction is planned to start in 2027.

What do you look forward to the most during your sabbatical?

I’m looking forward to connecting with other AI/deep learning researchers to get inspiration and ideas from them. I find it exciting that deep learning is interdisciplinary so that similar techniques work in completely different research fields. This provides excellent opportunities to learn from each other.

Autonomous Research Platform for Battery Electrolyte Discovery

Erik Berg, The Department of Chemistry, Ångström

Rechargeable batteries are a key technology to buffer an hourly alternating energy influx from wind/solar power and to enable distributed, flexible, and effective energy services. Batteries mainly consist of two classes of materials, namely electrodes and electrolytes. Most research today is focused on electrolytes, but progress is slow, primarily because the lack of

fundamental understanding and inherent complexities of these materials. The goal of my research project is therefore to accelerate the electrolyte discovery process by coupling data-driven methodologies (AI/ML) with a high-throughput robotic electrolyte evaluation platform developed in my lab.

What do you look forward to the most during your sabbatical?

Find inspiration, time, collegiality and support to integrate AI based methodologies into my research and educational activities.

Ultrafast X-ray Imaging of Heterogeneous and Dynamic Samples with Deep Learning

Filipe Maia, The Department of Cell and Molecular Biology

Since the invention of the first microscopes and telescopes imaging advances have helped drive scientific discoveries. Imaging fragile biological particles at the nanoscale and capturing their ultrafast dynamics is currently at the forefront of this quest. A single protein or virus cannot provide sufficient signal to obtain a good image, so multiple particles have to be averaged to boost that signal. Yet one must be careful such that the averaged particles are all similar otherwise we end up with a blurry image. This project aims to tackle this exact sorting problem in the field of ultrafast X-ray imaging, to explore how much heterogeneity can be tolerated, and try to understand and explain the dynamics we observe.

What do you look forward to the most during your sabbatical?

To get a better picture of how AI impacts other areas of science and, through discussions with the other fellows, learn from those experiences.

Assistive AI, Disability and Norms of Being Human

Amanda Lagerkvist, The Department of Informatics and Media

Today, a vast array of assistive and affordable AI technologies exists targeting people with disabilities. The purpose of this project is explore what norms about the natures of human subjects, communicability and bodies that can be detected in current AI developments and designs, as well as through how people with disabilities (PWD) appropriate the tools in their daily lives. Centering on existential and therapeutic chatbots as well as other assistive forms of AI (apps), AI Labs in Sweden, Switzerland and Germany working on language models will be studied as well as several patient organizations and NGOs in Sweden. The central research problem to be explored concerns how and if dis/ability plays into how AI is developing. Drawing on critical and historical research on the relations between disability and technology, the project explores how and if the two are potentially co-constituted in this context. With a particular emphasis on NLP innovations and applications and on how PWD appropriate these technologies in their everyday lives, the aim is to answer the following main research question: What norms about what it means to be human – and about embodiment and personhood – are expressed shaped, reproduced and challenged by those who are involved in developing assistive AI and among those for whom their everyday existence depends on and is pervaded by these technologies?

What do you look forward to the most during your sabbatical?

Focused time for both fieldwork and writing as well as the opportunity for new exchanges with AI experts in other fields and disciplines.

Addressing the Replication Crisis in Artificial Intelligence

David Johnson, The Department of Informatics and Media

Replicating others' research is crucial to building trust in published research results so that the artefacts generated from research (e.g., data, software, models, methods, techniques) can also be reused. The so-called “replication crisis” encapsulates qualities that are desirable for scientific research: replicability, reproducibility, and reusability, and there is much meta-scientific research on raising the bar of these qualities across research disciplines.

The field of AI is not immune to the replication crisis, and thus my project will focus on studying the research practices of AI research. Specifically, I will conduct an empirical analysis of AI-related research and replication practices to determine best practices that will be operationalised in a research replicability toolkit. My aim is that this project's successful implementation will increase robustness and trust in published AI-related research results.

To do this, I will first study research practices within a research group in the disciplinary domain of Medicine and Pharmacy at Uppsala University, which uses AI extensively. I also seek AI research groups to study beyond this domain and the scope of our university.

What do you look forward to the most during your sabbatical?

The nature of this project means that I will directly study others’ AI research. This will require connecting with AI researchers within and outside of Uppsala University, and I hope to build an extensive collaboration network to support the project.

But just as important for this project is disseminating the results (the recommendations for AI research best practices), and I am looking forward to connecting the research to teaching and outreach activities.

Tidigare forskningsprojekt

Här kan du läsa mer om forskningsprojekten inom artificiell intelligens och maskininlärning som beviljats medel tidigare år.

AI-lumner, projekten som fått medel 2022

AI-lumner, projekten som fått medel 2021