Research projects within AI4Research 2021

Name: Darek M. Haftor and Sandra Bergman

Project title: Harnessing the benefits of human-AI interactions in unstructured decision making: the quest for performance-conditioning factors

Department: Department of Informatics and Media

Area of research: Information systems

Project summary

This research proposal focuses on the performance benefits from the use of artificial intelligence (AI). Our focus is on the identification of factors that condition decision-making performance when humans and AI work together in unstructured decision situations. This documents starts with an introduction into the topic and the research direction assumed, which is followed with the specification of objectives for the AI4 Sabbatical work.

Research Project Focus

Even though firms and government organisations are increasing their investment in various kinds of AI systems, there is a lack of knowledge of how to ensure that the use of AI applications is beneficial. The intended outcomes of the here proposed research are expected to offer guidance on the use of AI in complex unstructured decision-making situations.

Uses of AI have found their way into domains such as logistics, industrial operations and areas of services including banking, finance and insurance. Until recently, the use of AI primarily targeted structured situations. In such situations, work processes and decision making are repetitive and well understood, with notable examples provided by robotics in car manufacturing and automation in the processing of insurance claims and mortgage applications. Such routinized decision making is characterised by unambiguous outputs, inputs and rules of the decision-making process. These characteristics make it possible to automate structured work processes and the associated decisions and thus replace human decision making with decisions by machines.

In addition to this established yet restricted use of AI in structured situations, recent attempts have been made to use AI to support, or augment, human decision making in unstructured situations. Real-life examples include Unilever’s talent acquisition process, Netflix’s decisions on movie plots, directors and actors, and Pfizer’s drug discovery and development activities. Unstructured decision making is characterised by ambiguous and/or changing objectives and decision criteria, with human judgement required to complete the decision-making process.

Three influential books on the use of AI for decision making provide normative guidance on how AI can improve organisational performance: Brynjolfsson and McAfee (2014), Daugherty and Wilson (2018), and Davenport and Kirby (2016). These books collectively suggest that the use of AI in augmented unstructured decision making benefits organisations more than mere automation of structured processes. Consequently, Microsoft will “build intelligence that augments human abilities and experiences. Ultimately, it’s not going to be about human vs. machine”. Similarly, in the preamble to its AI guidelines, Deutsche Telekom states that “AI is intended to extend and complement human abilities rather than lessen or restrict them”, whereas IBM affirms that “the purpose of AI and cognitive systems developed and applied by the IBM company is to augment human intelligence”.

Whilst all these intentions are well meant, extrapolations based on isolated lab studies are not satisfactory or rigorous enough. There is currently no conclusive evidence from real-life settings that AI-augmented decision-making processes perform better than unstructured decisions by humans alone. We know little about the contingent factors that condition the success of AI-augmented decisions. Therefore, the key objectives of the intended empirical research are (a) to provide rigorous empirical evidence that AI-augmented decision making in unstructured situations outperforms human decision making, and (b) to identify the factors that condition the success of AI-augmented decision making. Achieving these objectives will extend the currently limited knowledge base in this domain and will provide important normative guidance for organisations regarding the use of AI applications.

The here specified research direction builds on our recent project that received the 2019 Best Paper award from the “Strategic Management Society” for its identification of counterintuitive performance in AI-augmented decision making in unstructured situations. That PostNord co-sponsored study rests on a quasi-natural experiment, where a group of sales representatives in a firm performed significantly better in their sales decision making when supported by an AI-powered sales support system designed to provide a fit of several complementary factors. That fit was achieved by the matching of a decision maker’s cognitive style with two factors: the information representation mode offered by the AI system and the mode of work process conduct through which a given individual executes work tasks, uses the AI system and makes decisions. All this was compared with three control groups. Two groups used the same AI-powered system, without ensuring a fit with these complementary factors, whilst the other group used an older decision-support system without AI. The results show that the group with complementary fit performed best. Surprisingly, however, the results also show that the group without the AI-powered system performed better than the two groups with the AI-powered system unmatched to the complementary factors.

Overall, that study shows that AI-augmented decision-making performance is contingent upon a set of factors that must complement each other to generate the desired outcome. Crucially, this finding shows that any organisation – business, health or government – that invests in a wrongly configured use of an AI system may actually be worse off than not making any such investment and continuing with business as usual. Despite breaking new ground, that study has some limitations because it disregards two well-established factors that condition an individual’s decision performance (i.e. decision-maker’s motivation and decision authority span). My AI4Research Sabbatical intends to prepare for an empirical study that targets that gap.

Aim & Objectives for the AI4Research Sabbatical

Given the above, the overall aim for my AI4Research Sabbatical is to prepare for a successful execution of the above outlined research direction. That aim is to be achieved through realization of the following sabbatical objectives:

1. To have exposed the here assumed research direction to the AI4Research Fellows
2. To have established a researcher network within the AI4 context
3. To have established a research team for a joint research proposal
4. To have established cooperation with one or more industrial firms that commit to a participation in an empirical study
5. To have developed a full project proposal for empirical filed study that is to investigate the use of AI augmented decision making in unstructured situations
6. To have submitted the formulated project proposal to one or several funding agencies
7. To have developed detailed measurement instruments for the collection of data in execution of the proposed empirical study

The here assumed research focus involves theoretical components from several disciplines such as psychology, social-psychology, sociology, information systems and AI, and managerial economics. Uppsala University’s AI4Research initiative, with its Fellows from various disciplines, represents therefore an important intellectual environment for the formulation of a successful research project, as proposed here. The AI4Research Fellows may inform the present initiative (i) about additional theoretical components of relevance for this research direction, (ii) about how to link the various theoretical factors into a coherent framework, and then (iii) about how to practically utilise those theoretical components in an empirical study. The benefit of the AI4Research initiative may also have the other direction, as the research proposal outlined here will inform the Research Fellows about the state of the art in AI research that focuses the exploration of operational mechanisms which generate value creation from the use of AI systems.

Beside myself, this application includes also Sandra Bergman, a doctoral student at the Department of Informatics and Media. Her research is co-funded by Department and the Swedish National Research School on Management and Information Technology. Her interest focuses particularly on the conduct of decision-making processes in virtual teams, where at least one decision-actor is an AI-application, such as Google’s current use of its AI system Deep Mind to help scientists spread globally, working together to understand the devastating corona virus

Name: Anders Isaksson

Project title: Development of AI tools for improved analysis of tissue section images from cancers in a cross-disciplinary environment

Department: Department of Medical Sciences

Area of research: Molecular oncology

Project summary

In a pilot project funded by Cancerfonden we address the question of why some endometrial carcinomas appear to evade the anti-tumor response launched by the immune system. We test a number of specific hypotheses by searching for associations between molecular/spatial patterns in tissue sections and patient outcome. In order to reach this goal we train and apply deep learning algorithms to identify different spatial cell patterns in Hematoxylin & Eosin (H&E) stained tissue sections. As an example we identify the pattern of Tumor Infiltrating Lymphocytes (TILs).

Using the unique data sets available to us, as well as the outcome of our automated image analysis and TIL identification, we detect patterns associated to prognosis. We also search for novel molecular mechanisms underlying these patterns. In summary this program aims at providing improved prognostication with more certain and informative prognoses for patients and clinicians.

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

I am looking forward to discussing the common challenges we face while developing and applying AI methods to new applications.

Name: Åsa Johansson

Project title: AI tools to predict risk of common diseases using genetic data

Department: Immunology, Genetics and Pathology

Area of research: Human Genomics and Molecular Epidemiology

Project summary

In this project we will use Artificial Intelligence tools with genetic data, for prediction modelling of disease risk. During the last 15 years, we have performed genome-wide association studies (GWAS), and identified thousands of genetic variants to be associated with common diseases and disorders, such as obesity, myocardial infarction, allergies and asthma. Most common disease are complex, meaning that many (hundreds or thousands) of genetic variants influence the disease risk, together with other exposures, such as lifestyle and environmental factors. Methods available today are poor in accurately prediction individuals of high disease risk, predominantly due to the extremely large number of genetic variants in the human genome, the large number with an effect on each disease, the low effect by each such variants, and possible interactions between genetic variants. Our hypothesis is that by using AI, we can increase our ability to identify individuals that have a high genetic risk of developing a common disease

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

Human genomics is the fastest growing field with regards to data production and well suited for applying AI tools. However, there is limited experience in our research environment in using AI. I expect this sabbatical should lead to valuable interactions with AI researchers in other fields to share experiences and ideas, but I also hope for more persistent interdisciplinary collaborations.

Name: Matteo Magnani

Project title: Methods for the analysis of spreading phenomena in networks, with a focus on the online spreading of political ideas through visual content

Department: Information Technology

Area of research: Computational social science, data science, network analysis

Project summary

Images are a powerful means of sharing information and have been shown to increase the probability of information spreading and user engagement. In particular, images can be used to spread political messages (for example, for or against climate change mitigating actions), contributing to the definition of the public agenda. Current methods to study political communication through visual content rely on the manual inspection of selected pictures by domain experts, often in print news media. This is an approach that cannot scale to the complex contemporary communicative ecology made of multiple interconnected social and traditional media. This project will address the following research questions: How can we use AI methods rooted in network analysis to extend our ability to map and interpret online visual narratives in depth, at scale and across multiple social networks? What visual narratives around climate change exist and how can they be characterised with respect to their visibility, engagement, polarisation, and temporal evolution?

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

Having the time to set up this project, which has been developed together with Alexandra Segerberg (also a selected AI4Research researcher), identifying new areas where my research results can be applied, identifying new problems to solve, and also starting a broader cross-faculty initiative on the usage of AI methods in the social sciences.

Name: Carl Nettelblad

Project title: A deep learning Swiss Army knife for genomics and multi-omics

Department: Department of Information Technology

Area of research: Scientific Computing

Project summary

Neural networks have proven immensely successful in analysing image, video, and text data. So far, success has been more limited for genomic data. Together with my PhD student, I have created new adapted networks for genomic data. So far, we have applied those to questions in population genetics, showing how they can learn to identify the geographic origin of individuals with greater accuracy than the prevailing method of principal-component analysis (PCA). However, what’s most exciting about these networks is that they are general.

In conventional bioinformatics, each data type and problem formulation requires fully unique algorithms, but again, the examples from image analysis show how transfer learning can be used to adapt a network to a new problem. During the year, we will explore how our foundational network design can be adapted to address different genomic questions and even include proteomic and transcriptomic data.

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

I look forward to the new research ideas and renewed approaches to existing ones that will be possible in a truly cross-disciplinary environment.

Name: Alexandra Segerberg

Project title: Methods for the analysis of spreading phenomena in networks, with a focus on the online spreading of political ideas through visual content

Department: Department of Government

Area of research: Digital media and politics

Project summary

Images are a powerful means of sharing information and have been shown to increase the probability of information spreading and user engagement. In particular, images can be used to spread political messages (for example, for or against climate change mitigating actions), contributing to the definition of the public agenda. Current methods to study political communication through visual content rely on the manual inspection of selected pictures by domain experts, often in print news media. This is an approach that cannot scale to the complex contemporary communicative ecology made of multiple interconnected social and traditional media. This project will address the following research questions: How can we use AI methods rooted in network analysis to extend our ability to map and interpret online visual narratives in depth, at scale and across multiple social networks? What visual narratives around climate change exist and how can they be characterised with respect to their visibility, engagement, polarisation, and temporal evolution?

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

I’m looking forward to two things. The first is the opportunity to develop this project together with Matteo Magnani (also an AI4Research scholar). Ours is a cross-faculty collaboration, so getting the chance work at the same place for a sustained amount of time will be invaluable for grounding a shared base to build on. The other is the contact with other projects in and around the research hub. An underlying aim of the collaboration is to open up not just possibilities, but also challenges and implications of implementing AI and deep learning methodologies for social science purposes. The chance to have a broad and informal discussion across projects and disciplines will enrich this process.

Name: Erik Zackrisson

Project title: Machine Learning in Astronomy

Department: Department of Physics and Astronomy

Area of research: Astronomy

Project Summary

The night sky is like a time machine – the light from the most distant astronomical light sources has been travelling for more than 13 billion years to reach us. By studying such objects, we can learn more about the formation of the first generations of galaxies, at a time when the Universe was just a few percent of its current age. In 2021, the James Webb Space Telescope will open a new window on this epoch in the history of the Universe, but since galaxies at extreme distances also appear very faint, the data will inevitably be noisy. Through the use of machine learning techniques, we hope to increase the science return of these new observations of the early Universe.

While the study of the very distant Universe can be said to be data-starved, the opposite is true for the nearby Universe. Thanks to large sky surveys, we have access to rich sets of data for billions of light sources in our own cosmic backyard. Here, machine learning can help us find anomalous objects, potentially leading to the discovery of previously unknown astronomical phenomena.

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

Focusing on the development of new research methods and getting influences from other fields.

Name: Fiona Skerman

Area of research: Random graphs, network theory, probability

Department: Department of Mathematics

Project title: Fundamental limits of learning in community structure

Project summary

A key challenge in AI and machine learning is to find limits to what one can 'learn' about a data set: what can be realistically inferred about noisy data. Here we represent the data as a network and ask whether we can detect the difference between two networks, one of which contains communities which represent structure in the data. This project is to investigate in what circumstances it is possible for any learning model to detect the communities in non-perfect data.

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

The opportunity to have dedicated time to think on this project and to interact with my visitors and other sabbatical holders. I think it's important to be always talking to researchers using network data to inform their research to ensure we study the network properties of interest and that we consider useful models of the errors in network data.

Name: Ida-Maria Sintorn

Title of your project: Why doesn’t it work in reality?
– Bridging the gap between curated proof of concept tests and real world deployment of biomedical image based deep learning

Department: Department of Information Technology

Area of research: Computerised Image Processing

Project summary

Deep learning has lifted image processing and analysis to a whole new level, especially so for computer vision applications based on the enormous amounts of accessible images on internet. Biomedical applications follow but face a different scenario with less available training images, impact of erroneous results, trustworthiness and credibility. This project will focus on increasing the understanding of how to best incorporate application domain expertise to mitigate some shortcomings hindering the deployment of deep learning solutions in real-world clinical scenarios. The aim of the project is to develop interactive verification and improvement approaches to increase the credibility and trustworthiness of image based deep learning in biomedical/clinical applications. More specifically: 1) to explore strategies to identify rare or unexpected classes not encountered in the training set, and 2) to develop methods to interactively incorporate expert feedback regarding what features are important/false.

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

The multi- and interdisciplinary setting.