Artificial intelligence has created unexpected opportunities and new trains of thought in many fields of research. Physicist Daniel Midtvedt and statistician Rebecka Jörnsten are two of the Faculty’s researchers who use AI in their work. Both believe that AI is going to revolutionise many fields, but that it is important to be aware of its limitations.
dazzling laser beams cast a green glow over the darkened lab in the Physics Building. It’s where research is being conducted using a holographic microscope, which makes it possible to see how all the components in a sample interact. For Daniel Midtvedt, Associate Senior Lecturer at the Department of Physics, AI is a tool for extracting more – and more reliable – information from microscopy images. This information can be useful in everything from materials development to a better understanding of how cells work.
But explaining what AI really involves is no easy task, even for an associate senior lecturer. Daniel Midtvedt summarises AI as “a system of smart algorithms or networks that can learn to handle certain specific situations in order to achieve a goal”. He likens it to a black box that you feed with mathematical models and data. This data is then processed by computers, and a result comes out on the other side.
“The strength of AI-based networks is that they can break down extremely large volumes of data into more manageable material. Because we let the computer draw the conclusions based on the data we give it, bypassing whatever preconceptions we scientists might have about the results, and this makes it a bit more objective,” he says.
he was attracted to AI when he realised that traditional methods could not provide the information he needed for his research.
“AI makes it possible to start from complex microscopy images with an extremely large amount of information, and extract precisely the information we want. Without AI, we would not have been able to monitor and quantify the physical properties of nanoparticles in cells. But with this technology, we will now be able to study how viruses and cells interact, and what steers the process in various diseases like Alzheimer’s. The hope is that these technologies can eventually be used in medical research, for example to develop new drugs,” he says.
In many of his projects, Daniel Midtvedt collaborates with the business sector as well as other research groups at the University of Gothenburg, including some external to the Faculty of Science.
“A lot of people are interested in AI and that interest is growing. The multidisciplinary nature of the field is reflected in the many different research projects in progress. And it’s clear that there is demand in the business sector for AI-related research,” he says.
Rebecka Jörnsten, Professor of Biostatistics and Applied Statistics, agrees. She works at the Department of Mathematical Sciences and participates in a number of interdisciplinary AI projects. For example, she collaborates with biologists and medical researchers with the aim of developing individually tailored cancer treatments.
According to Rebecka Jörnsten, great interest in AI has meant a lot for mathematics research.
“AI has opened up opportunities for our researchers to take a leap and try something new, even if it’s outside their comfort zone. This has led to exciting cross-fertilisations and spin-off effects. For example, many mathematics researches are looking at classical statistics in entirely new ways,” she says.
Rebecka Jörnsten’s biggest research field concerns basic research into deep learning models. It is an area of AI in which neural networks are trained to solve problems in many small steps. With a sufficiently large set of training data, these networks can ‘learn’ to solve problems that were impossible with the calculation methods used in the past. The problem that Rebecka Jörnsten and her doctoral students are currently tackling is that the learning process must not be too fast.
“If the learning process is too fast, these networks, or models, get hung up on noise, meaning things that you don’t want to model. We are looking at ways to slow down the learning process, including statistical regularisation, where the computer learns to recognise errors and ignore them. The big advantage of a slower learning process is that it allows us to train very complex models using even fairly small data sets,” she says.
At first sight, explaining the importance of basic research and methodology development might appear to be more difficult than explaining applied research. But Rebecka Jörnsten has a clear message.
“All methods consist of different building blocks. Having a good building block can play a crucial role, and we’ve created a very good building block now! Anyone who uses neural network training in their research can benefit from what we are doing, and it may ultimately lead to many applications that will benefit society,” she says.
So what then are the challenges and limitations in AI? There are a great number, according to both Rebecka Jörnsten and Daniel Midtvedt, who also stress the importance of being aware of them. One problem is that it can be difficult to deduce what flexible, self-learning AI models are basing their conclusions on.
“I think we need some kind of framework or standardised protocol for validating results from AI models, to know when and to what extent we can trust them,” says Midtvedt.
Rebecka Jörnsten highlights the ethical aspects.
“Just because we can create a certain system, it doesn’t mean we necessarily should. For example, there are many AI-based, practical everyday solutions that are great, and AI can be used to improve road safety or reduce queues in healthcare. But I think we should be careful about using AI systems in human interactions. A model can be 99% right, but that one per cent that it gets wrong can be entirely unacceptable in the wrong setting,” she says, stressing the importance of thinking about how the data is collected.
“It’s important to understand that data quality must be the crucial factor. If you give an AI model ‘bad’ data, such as unrepresentative or unbalanced data, then you will also get a bad model.”
But both of these scientists also believe that AI will continue to create new research opportunities in many different fields, which can lead to great benefits for the community at large.
“I think AI will be used more and more to quantify systems and analyse complex data sets. In my field, for example, I imagine that AI will be used to assist medical professionals in their decision-making.”
A little AI glossary
What does a ‘neural network’ mean and what is the difference between AI and machine learning? Daniel Midtvedt clarifies some artificial intelligence terms.
“AI can be described as a system of smart algorithms that process large quantities of data on fast computers. These algorithms can apprehend and learn from their environment, as well as handling and solving problems to achieve a specific objective. In our case, it’s interpreting a microscopy image, but it might just as well be about recognising a human on the pavement as part of the development of self-driving cars.”
“A subcategory of AI, or a tool for achieving AI. In machine learning, computers can be ‘trained’ using various datasets to detect and ‘learn’ rules for solving a problem, without the computers having been programmed with rules for that particular problem. In other words, machine learning helps machines to learn to understand a problem completely on their own.”
“Neural networks are self-learning algorithms that try to mimic the functions of biological neural networks – like the neurones in our brains. Neural networks, like other AI systems, create their own rules using machine learning. They can handle very complex information and are trained to draw conclusions themselves.”
Deep learning or deep neural networks are also part of the field of machine learning. The fundamental idea is training a very deep (of many, many layers) neural network so that it can solve a problem in many small steps. With a sufficiently large set of training data, neural networks can solve problems that were impossible with the calculation methods used in the past.
Three voices on...
...developments in AI. What are the most important areas in which AI can make a difference in the future? What are the biggest AI challenges from your point of view?
Senior Researcher in Machine Learning at RISE
“Some of the most important applications for AI in the future will be in environmental and health areas. We can measure courses of events in greater and greater detail, which gives us a better basis for predicting phenomena such as floods and droughts, and data that can be used to train data-driven models. Climate change means that we need to be prepared for more extreme weather events, but we also need tools to ameliorate this process.
“AI-based solutions need to be more capable, more accessible, and more reliable. We need to make it easier to apply AI in areas where it was not possible to do so previously. This includes methods for collecting and processing data, and methods for determining what technologies are suitable for which purposes. AI technology must be able to estimate and communicate the reliability in a decision if users are going to trust it.”
Head of Strategic Initiatives, AI Sweden
“AI is already useful in almost every field and industry. One area that has advanced rapidly is the development of large-scale language models that understand and generate human language. I also believe AI is going to contribute to breakthroughs in many fields of research. For example, there are now AI systems that predict the 3D structure of proteins, and image recognition systems that have led to new insights into bird behaviour.
“Questions related to data management, data access and data quality are crucial challenges that must be solved, and there are many legal challenges too. I also see challenges in operationalising AI on a large scale, in other words moving from limited test environments to applying AI on all fronts across an entire organisation. Few organisations today have the capacity and knowledge to do this. But we will only see the really big benefits when they do.”
Head of Research, Technology & Collaboration at Zenseact
“People who use AI as a tool are going to replace those who don’t. AI is very good at perceiving and recognising patterns and never gets tired or loses focus. AI is going to be particularly helpful in traffic in the area of perception, where a car needs to recognise patterns in complex traffic situations in differing light and weather conditions, but also for trajectory planning in complex traffic situations.
“One major challenge is to prove that AI technology is sufficiently reliable, meaning that it generalises well enough in all situations that can arise in traffic, for example. It’s important to be successful in developing high quality models that minimise privacy risks, but also to reduce the energy consumption for training the models. Finding a balance between what is technically possible and what is legally possible is also vital.”