Awarded Research Projects in Artificial Intelligence and Machine Learning

AI-lumni – projects 2021

Projects 2022

The AI Risk Narrative – a transdisciplinary inquiry

Magnus Strand, Department of Business Studies

Magnus Strand
Magnus Strand.
Photo: Mikael Wallerstedt.

In 2020, the research environment in law and business studies which I am leading acquired two major grants for AI research from WASP-HS. Both these projects (and a third which is pending) connect to theoretical issues connected to the concept of risk, and to what I refer to as the AI risk narrative. There is a need for close analysis of how this risk narrative is being used in legal argumentation (by legislators, authorities and courts) and this is what I intend to do at AI4Research. Such an analysis can contribute to self-reflection in legal discourse and, hopefully, to an informed and professionalised on risks and opportunities in AI applications. In extension, the analysis can be helpful for regulators.

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

Besides carrying out my project, I am looking forward to learning more on what AI studies are ongoing in Uppsala University. Hopefully this will enable new collaborations between my own and other research groups in the University.

Sparsity Driven Statistical Learning in High Dimensions

Tatiana Pavlenko, Department of Statistics

Tatiana Pavlenko.
Tatiana Pavlenko.
Photo: Sergiy Kupriyenko

The long-term relevance and success of modern Artificial Intelligence (AI) systems and technologies depend critically on theoretical foundations, and in particular on  concepts and theorems of statistical learning theory that provide a deep understanding of recent fast progress in AI.  One of the topics  of fundamental importance for the modern  theory of statistical learning is sparse recovery,  i.e. the ability to detect sparse informative signals from noisy, high-dimensional data.

In practical applications sparsity is often combined with weakness, meaning that individual effects are not sufficiently strong to be detectable using conventional statistical procedures such as e.g. large scale multiple testing. In view of modern computational techniques, statistical models that can combine sparsity and weakness with high dimensionality are tractable and, in fact attractive in practice, but efficient statistical learning in such models poses theoretically challenging problems. In a high-dimensional scenario, the classical Fisher-Le Cam theory of the optimality of ML estimators and Bayesian posterior inference do not apply, which explains a flurry of recent activities in providing full-fledged inferential and algorithmic capabilities of these models. The current project is aligned with these research directions and aims at the methodological developments in the theory of sparse signal detection and its applications to the complex, high-dimensional data sets arising in modern science and engineering.

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

The opportunity to have time to focus on my research ideas which is very important since academic life can be very intense. Due to pervasiveness of high-dimensionality in modern data analysis, I hope that the project's activity will attract attention from AI4Research Fellows working in diverse applied fields. During this sabbatical I hope to establish a cross-disciplinary research network within AI4Research where new methodologies of large-scale statistical machine learning will be adopted and further developed to provide useful insights into the practical aspects of AI applications. 

Studies of peptide drug solubility and permeability from physics-based simulations and artificial intelligence

Per Larsson, Institutionen för farmaci

Per Larsson.
Per Larsson.
Photo: Mikael Wallerstedt.

We are developing a computational framework that will allow modeling of key processes in the intestine so that the limitations and variability in bioavailability for peptide drugs can be understood, even on an individual level. One of the principal limitations here is quantitatively accurate in silico predictions of drug solubility and permeability.  By combining physics-based simulation with machine learning algorithms, we hope to leverage the best of both these worlds so that delivery systems for peptides can be designed specifically for different individuals. In the end, this will reduce variability and increase bioavailability of orally administered peptide drugs. 

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

To have the opportunity to do my own research in the interface between simulation and AI / ML, and, perhaps above all, the chance to discuss my research with experts from other areas within the university and in AI, to find synergies.

Predicting peripartum depression using a smartphone application and digital phenotyping

Fotis Papadpoulos, Department of Medical Sciences

Depression during pregnancy or after childbirth (peripartum depression (PPD)) is a common (ca 10%), serious and potentially life-threatening disorder with high societal costs. Preventive efforts targeting PPD are thus of paramount importance, especially among high-risk groups, but our current ability to predict PPD is poor.  We propose a novel, transdisciplinary project to design, develop and evaluate effective and user-centered methods for the prediction of PPD. We will use self-reported measures, voice recordings and digital phenotyping (passively-registered data) through the Mom2b mobile application to predict the development of PPD in the third pregnancy trimester, as well as during the early and late postpartum period. This will be achieved through implementation of state-of-the-art machine learning methods. Finally, we will explore the experiences, attitudes and possible concerns of the participating women while using the Mom2b smartphone application, and their suggestions for further improvement in order to assure a user-centered approach in Mom2b’s further development efforts.

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

To meet other researchers applying machine learning in their projects, learning from each other, so that I can develop my own skills in this complex promising field.

Using AI to detect and characterise atmospheres of habitable exoplanets

Oleg Kochukhov, Department of Physics and Astronomy

Oleg Kochukov.
Oleg Kochukov.
Photo: Mikael Wallerstedt.

Investigation of planets orbiting stars other than the Sun (exoplanets) is one of the most exciting and active directions of modern astrophysical research. By the beginning of 2021 more than 4000 exoplanets were discovered and catalogued. The bulk of these discoveries came from two indirect methods – observation of the radial velocity variation of a host star caused by the orbiting planet and detection of dimming of the stellar light when a planet passes in front of its host star. Using a combination of these techniques, astronomers learned that our nearest neighbours and many other stars in the Galaxy have planetary systems. The next major challenge of exoplanet research is to advance from a detection to a detailed characterisation of exoplanets. This requires measuring pressure and temperature in exoplanet atmospheres and determining their chemical composition. Of particular interest are potentially habitable exoplanets – smaller rocky worlds with the atmospheres similar to Earth.

The fundamental basis of the exoplanet characterisation research is detection and analysis of the spectra of exoplanet atmospheres – a daunting problem that is being attacked by many researchers worldwide using the largest ground-based and space telescopes. But even the most powerful telescopes currently available cannot discern an exoplanet spectral signature without sophisticated and non-trivial data processing. The corresponding computational techniques and computer codes are not widely available and severely lacking in scope and efficiency. New numerical methods are urgently needed to make best use of the data collected by costly astronomical experiments. In this project I plan to harness the power of AI to address the key problem of the cutting edge research on exoplanets: how to detect and interpret spectral signature of an exoplanet’s atmosphere?

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

I am looking forward to expand my knowledge in the field of AI, learn from other fields, establish cross-disciplinary collaborations and apply new research methods in astrophysics.

From Data to Molecular Biophysics 

David van der Spoel, Department of Cell and Molecular Biology

David van der Spoel.
David van der Spoel.
Photo: Mikael Wallerstedt.

Despite five decades of work there is no accurate mathematical relations to predict the energy of a molecule from its coordinates. Here, I will use large quantum chemical libraries in conjunction with experimental data on physicochemical properties to discover such relations. I have previously used Bayesian Monte Carlo to approach this problem and within AI4Research I hope to get a deeper knowledge of this and other machine learning tools.

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

To discuss science with my peers at AI4research, to learn new techniques. 

Many faces of science: AI-driven methods for multimodal image analysis towards interpretable information fusion

Natasa Sladoje, Department of Information Technology

Technology is helping life sciences to observe what previously was hard to even imagine. Powerful and complex imaging techniques reveal a variety of properties of a specimen – morphology, chemistry, dynamics, function – however, most often only one such property at a time. For full understanding, different complementary techniques have to be combined. The aim of this project is to explore and develop new approaches for multimodal deep information integration and analysis, based on learned cross-modal disentangled representations and Explainable Artificial Intelligence (XAI). We will contribute to the development of theoretically well founded, modality- and application-agnostic methods which enable reliable, interpretable and trustworthy utilization and analysis of information-rich multimodal image data.  We plan to apply our developed methods in a number of scenarios, primarily within biomedicine, and to evaluate their potential to advance everyday health care, but also to stimulate further research and knowledge discovery.

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

I am looking forward to the discussions with other AI4Research fellows, which will, as I expect, inspire exploration of novel ideas and enable their evaluation and refinement. I will be happy to strengthen collaborations and find synergies by exploring possible application scenarios in different scientific contexts where multimodal image data is used. Last, but not least, I am looking forward to the focused research time that this sabbatical enables.

Machine Learning in Theoretical Chemistry

Roland Lindh, Department of Chemistry – BMC

Roland Lindh.
Roland Lindh.

Summary of your project: Optimization problems: we have in the past demonstrated superior behavior of Gaussian Process Regression (GPR) in connection with molecular geometry optimizations . We would like to explore this further. First we like to extend the use of GPR to the procedure of finding so-called conical intersections between multiple potential energy surfaces. This mean extending the surrogate model from reproduce a single variable to that of the matrix element of the Hamiltonian for at least 2 states. We would also like to explore and develop “light” versions of gradient-enhanced kriging (GEK) in association with the optimization of the parameters of standard wave function models. This since GEK suffer from poor scaling as the number of parameters increase – wave function models could easily contain 102-109 parameters – the inversion of the covariance matrix becomes prohibitive.

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

To work in an interdisciplinary environment.

Explainable AI – What and Why, not just Where
Towards more useful explanations of deep learning-based image classification

Joakim Lindblad, Department of Information Technology

Joakim Lindblad.
Joakim Lindblad.
Photo: Marina Matić.

With the growing prevalence of AI and convolutional neural networks (CNNs) to support decision making throughout society, there is an urgent demand to explain what their decisions are based on. This has led to the creation of the rapidly growing field of explainable AI (XAI). Several techniques have been developed to shed light on how the decisions of AI systems are made. However, the vast majority of such methods are offering explanations of the type where in the image, or possibly when in the video, the important factors are appearing. These saliency, activation, attention, or attribution methods highlight where the network is “looking” when reaching a decision, but they do not answer what the network finds as important at these positions. Is it colour, shape, texture? Neither do they provide information on how a combination of different properties leads to the output of the network. Is it the fact that the observed cell is unusually large, in combination with having a coarse texture, which leads to it being classified as possibly malignant, or is it the combination of a thick nuclear envelope and the particular hue of the cytoplasm? This project aims to develop methods which can deliver these types of information, enabling improved understanding of decision making – essential for confident usage of AI for critical tasks, but also for better insight in the underlying phenomena – facilitating new knowledge discovery.

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

To find time to focus and delve deep into my research, and to discuss and exchange ideas and experiences with colleagues.

Machine Learning Accelerated Electrolyte Modelling and Design

Chao Zhang, Department of Chemistry - Ångström Laboratory

Chao Zhang.
Chao Zhang.

Electrolyte is an indispensable and critical component in any electrochemical storage devices. In this regard, my goal for the stay at AI4Research is two-fold: i) to initiate a new direction in my continuous effort on combining deep-learning and molecular simulation for modelling electrochemical systems; ii) to bring experimental and theoretical colleagues into the discussion for possible new collaborations.

What do you look forward to the most during your sabbatical? A new and stimulating environment at Carolina Rediviva.

AI-lumni – projects 2021

Nine research projects associated with artificial intelligence and machine learning have been awarded funding within AI4Research. They concern information systems, molecular oncology, human genomics, molecular epidemiology, computer science, network analysis, digital media, policy, astronomy, network theory, probability, and computer-assisted image analysis. You can read more about the research projects here.

Darek M. Haftor and Sandra Bergman, Department of Informatics and Media
 
Anders Isaksson, Department of Medical Sciences
 
Åsa Johansson, Department of Immunology, Genetics and Pathology
 
Matteo Magnani, Department of Information Technology
 
Carl Nettelblad, Department of Information Technology
 
Alexandra Segerberg, Department of Political Science
 
Erik Zackrisson, Department of Physics and Astronomy
 
Fiona Skerman, Department of Mathematics
 
Ida-Maria Sintorn, Department of Information Technology

Last modified: 2022-03-21