What can be found in a buzzing brain?

Even the most efficient computer “just works here”. Somebody must instruct it to use its computational power sensibly. Aapo Hyvärinen has instructed computers to sort data from brain imaging.

A building constructed by the ATR research institute near Kyoto measures the residents’ brain activity whenever they are at home. For Aapo Hyvärinen, a specialist in machine learning, these families provide a veritable cornucopia of data on human brain activity.

Functional neuroimaging has traditionally been performed in laboratory settings where the subjects must complete simple tasks for a few minutes while wearing sensors on their head. But this approach does not satisfy researchers.
“The global trend in brain imaging is to move away from laboratories to more natural conditions,” says Hyvärinen, who works as a professor of computer science at the University of Helsinki.

Regularities in the multitude

Moving away from laboratories and into natural situations leads to a huge increase in the amount of data, which also poses considerable requirements on computers. Although their computing power nowadays is considerable, it it is not enough. They must also be directed to use that power sensibly.

This is precisely the field – independent component analysis (ICA), which is a part of machine learning theory – in which Hyvärinen has conducted his pioneering research. For his cutting-edge research, he received an invitation to spend a year in Japan exploring the data generated in the Kyoto building.

ICA examines the observation data for interesting, peculiar components similarly to the way in which people concentrate on a single interlocutor during a buzzing cocktail party.
“Brain imaging data are one of ICA’s most comprehensible applications. We can find plenty of regularities in the multitude of data. These regularities are signals from small brain regions that function somewhat independently,” explains Hyvärinen.

Straight to the top

The basic dilemma of modern science is that even the best measuring device can seldom answer the research question directly. It simply produces some sort of data. “Finding regularities in the data is difficult,” notes Hyvärinen.
“Although you may understand the regularity on an intuitive level, formulating it into an exact mathematical theory and then into a computer algorithm that seeks that regularity often requires thousands of person-years of effort.”

Hyvärinen describes the beginning of his career (i.e., the doctoral dissertation that he began preparing in 1995) as “mathematical/theoretical dabbling”. His topic was independent component analysis.  

“At the time, people already understood that an increasing amount of measuring data was being produced in different fields and that the computing capacity of computers had increased. But the development of theories and algorithms that can interpret the data was in its infancy,” Hyvärinen recalls.

In his dissertation, Hyvärinen conceived an algorithm that could perform the required computations faster, but without compromising accuracy. His major innovation was to apply a fixed-point algorithm in independent component analysis. “Turning my theoretical idea into an algorithm eventually happened fast, in only a few months. And the end result was also quite simple: about a dozen lines of commands in a scientific computing environment,” Hyvärinen says.

When he completed his dissertation in 1997, applied research was already underway, and Hyvärinen’s algorithm proved more efficient than those of his competitors.

The algorithm became widely used and has been applied over the past decade or so to different data sets in various disciplines. It has found interesting components in, for example, space images and genetic data. The algorithm also continues to generate citations for Hyvärinen from around the world.

According to Hyvärinen, “It’s a bit annoying that I made my most cited discovery in my first year of postgraduate studies. I have since expanded the algorithm, but the expansions are not nearly as significant as the original algorithm.” 

Powered by green tea

After completing his dissertation, Hyvärinen’s career has focused on the brain. First, he applied his algorithm to constructing computer simulations of the human visual system, and since 2007 he has been developing algorithms for the interpretation of brain imaging data.

Now, thanks to the Japanese inhabitants of the Kyoto building, he is receiving an unparalleled amount of brain imaging data.  “Analysing the data requires not only computing capacity, but also a lot of green tea!” notes Hyvärinen.

Text: Antti Kivimäki