Samuel Kaski teaches machines to learn

Samuel Kaski teaches machines to learn

18 Oct 2018

A portrait of Teuvo Kohonen, Finnish Academician of Science, hangs in the entrance hall of the Department of Computer Science at the Otaniemi campus of Aalto University. And for good reason. Kohonen trained several generations of AI scientists during a time when very few in Finland even talked about artificial intelligence or machine learning. Kohonen was ahead of his time with his Self-Organising Maps.

One of Kohonen’s former students is Academy Professor Samuel Kaski, who is currently heading an initiative called Finnish Center for Artificial Intelligence (FCAI), which will pool together top AI research from across the country. “I had studied computer science and neuroscience for a few years when I got a summer job on Kohonen’s research team,” Kaski remembers.

“I had been completely unaware that there was something so interesting in Finland. I had done very well in my studies, but I was still amazed that I got a chance to join the team. Looking back, it was a hugely important stepping stone to get involved in top-level research so early. I had many interests, but doing research completely swept me away when I understood that it can be a proper job.”

Today, Kaski himself takes on promising students for his research projects as early as possible. “It’s a good practice that is ubiquitous in cutting-edge international research. Although I myself moved on to other lines of research soon after obtaining my PhD, I like to tell this story every so often in case there are young bright students in the audience. Maybe some of them get a spark to try out research.”

Tackling cancer with AI

Kaski’s interests lie in automatically learning models of the world, how such learning may help in understanding events and how the models may be technically applied for various other purposes.

“These models may be completely data-driven, or they may already contain knowledge on how things work. Usually, it’s a combination of these two. Further data are needed to learn the unknowns. And, we need an algorithm that is capable of fitting the model to the available data.”

To make the concept more tangible, we need only look at what Kaski has been up to recently: he has utilised machine learning for medical applications.

“A major question within genomics is how we can extract key bits of information from genomic data to aid disease modelling and better predict the effectiveness of treatments. The data contain information on gene function, metabolomics and cellular measurements. Take cancer samples, for example: we’ve been able to find out what is relevant in the datasets, and in their dependencies. This has improved the prognoses of what kinds of therapies are effective for each patient, analysed based on a specific tissue sample,” Kaski says.

Studying rare diseases is particularly difficult. Kaski has managed to develop methods to dig out key information in cases where there is not enough data on a new patient or on patients suffering from the same disease.

“The trickiest part is that although there’s a very small amount of essential information, there’s an insane amount of measurement data. In the case of genomics, there may be millions of potentially useful variables. Based on the data, we must be able to pick the variables that should be taken into account. That’s why we have combined datasets and focused on the interconnections between variables instead of individual data.”

Drug development faces a similar problem. The way in which medicines work is that the active substances target specific proteins. At the same time, they also influence other proteins whereby they may even cause harm.

“What we have is a giant matrix of proteins against drugs. AI makes it easier for us to predict what’s lacking from the matrix – especially in sections that could be of use in advancing pharmaceutical development,” Kaski says.

Much of medicine today is highly data-driven. More data and easier data collection facilitates entirely new ways of asking research questions.

“Traditional drug development is highly expensive and takes very long. At the same time, we’re seeing a rise in antibiotic resistance and decreases in the efficacy of many drugs. That is why the effectiveness of the drug development process must be improved for humans to come out on top in the race against pathogens.”

Human meets machine

Kaski often emphasises that machine learning, AI and the many applications of AI function best when they are harnessed to support humans. Here, too, medicine provides a concrete example: visiting the doctor.

“Machines can be extremely useful, for example, when physicians decide what measurements to take and when they make diagnoses and recommend treatments. Physicians are highly trained experts, but AI can provide them with a lot of additional information based on the millions of measurements in genomic data.”

This is where precision medicine comes into play. Precision medicine utilises even cellular-level data to tailor medical treatment to the individual characteristics of each patient.

“However, this won’t reduce the need for doctors; above all, patients will still choose their own treatment. After all, all forms of treatment come with both benefits and side effects,” Kaski says.

The terminator scenario

Artificial intelligence and its applications are tools that can be used for both good and bad. Kaski argues that the greatest threat is the same as with other forms of modern technology: that people will wield AI as a weapon against each other.

“AI may be a double-edged sword with great advantages and disadvantages. That’s why it’s crucial that we consider its ethical issues within the framework of a democratic society. We need scientists who will study the potential threats so that we can minimise the risks involved. But I still think we’re too hung up on the sci-fi scenarios where terminator cyborgs rise to take control. It is a risk in theory, but it’s still a much smaller risk than many others we’re facing in society today. I’d like to see more discussion about how to harness AI for more effective work and for solving the grand challenges we are facing,” Kaski says.

Artificial intelligence is also a societal issue. In any event, we are facing great changes as we enter a time when many traditional work tasks will become unnecessary. According to Kaski, very few jobs will actually disappear, but most will require some reskilling.

“It’s happened before! With the invention of writing, practically all jobs at the time had to adapt and people needed to adopt note-taking skills and information sharing. As for AI, the changes will happen much faster. The changes may make society more equal or more unequal, and our task as a democratic society is to choose from these two.”

Ideas come not only in the shower

Kaski has been – and evidently still is – interested in a wide variety of things, but he nevertheless urges people to focus on what is essential.

“My own field is method development. Its blessing and curse is that a good method is applicable to many different tasks. What this means is that I need to develop both the theory and practical applications that are in demand in very many fields.”

Kaski says that he has chosen the areas of application based on their potential for interesting collaborations. Lately such collaboration has flourished particularly within medicine, where genomics is offering up a great deal of data from which AI can produce exciting, new applications.

“It’s crucial to have a scientific community where scientists will bump into colleagues with completely novel yet compatible ideas. This is where the innovation happens. It’s true that people may get their best ideas while in the shower, but usually that has been preceded by discussions at the right time with the right people.”

Original text in Finnish and photos by Jari Mäkinen

Last modified 19 Oct 2018
Connect with us
Facebook Slideshare Twitter Youtube
SWITCHBOARD +358 295 335 000
REGISTRY +358 295 335 049
FAX +358 295 335 299
OFFICE HOURS Mon–Fri 8.00–16.15