Using real-life materials
A self-organising map starts off with a dataset, which is fed into the system in numerical form. Artificial neurons learn through repetition of an algorithm to independently classify the input’s different properties without knowing the correct answer beforehand.
Oja remembers how Kohonen emphasised that the algorithms should be tested as realistically as possible, using authentic materials. The first practical applications of the self-organising map focused on speech recognition. Sound spectre recordings of speech some ten milliseconds long were used as input.
Helsinki University of Technology’s Neural Network Research Centre, which was led by Kohonen, also developed the WebSom method that utilises the idea of the self-organising map for the purposes of organising and searching digital texts. It creates a visual representation from text material in which texts that are similar in terms of content are placed in each other’s proximity.
‘WebSom did a lot to make Kohonen a recognised name. It was one of the first really big neural network applications.’
Serious about the work
Over time, the self-organising map accumulated thousands of research citations as well as a large selection of practical applications. Kohonen’s invention was applied to analysing large data volumes in fields ranging from chemistry to communications technology and from medical science to the process industry.
Oja says Kohonen realised quite early that the global volume of data was about to undergo explosive growth. This was still a novel idea in the computer science of the early 1980s.
‘Neural networks are popular precisely because they make the handling of massive datasets possible.’
Oja remembers Kohonen as not only a brilliant researcher, but also a demanding personality who took scientific work very seriously indeed. His point of departure was that the research problem should always be on your mind.
‘If you’re engaged in scientific work, then do it seriously, Kohonen emphasised. Think about it all day, even in your sleep,’ Oja says.
Kohonen also observed these principles in his own life. He published his final textbook, which is crammed with scientific and technical calculation examples written in the MATLAB language, in 2014, the year he turned 80. The book focuses on using the self-organising map in the MATLAB computing environment.
Impact still felt
Deep learning neural networks and other AI methods surpassed Kohonen’s map after the turn of the millennium, but the SOM still pops up from time to time as a data organising and visualisation tool in research papers from a variety of fields.
The mark left by Kohonen is still felt in domestic science circles, says Professor and Finnish Center for Artificial Intelligence Director, Samuel Kaski, who studied the self-organising map in his doctoral thesis.
First and foremost, Kaski characterises Kohonen as a scientific innovator who had strong faith in his own views. At the same time, he demonstrated that Finland can produce pioneering research in the field of artificial intelligence as well.
‘This is precisely what I myself found enticing. I wanted to learn from a person who was engaged in pioneering work,’ Kaski says.
Retrospectively, it is easy to say that the success of Kohonen played a part in cementing the strong role of AI as a research field at Finnish universities.
‘Finland is now home to very significant AI research, also on the global scale.’
