Machine learning is a branch of artificial intelligence in which computers, without the help of humans, process and analyse large amounts of data in order to find patterns. Recent progress in this area has led to high-performing innovations in many areas, such as specialised healthcare, where computers have learnt to diagnose skin cancer .
“It’s about teaching computers to see patterns that are useful for us”, says Dylan Mäenpää, a former student at the Department of Computer and Information Science at Linköping University.
In June 2021, he presented his degree project about how the speed and security of machine learning can be improved further. He has now received a Christer Gilén scholarship, worth SEK 10,000, as a reward for his work.
“It’s always nice to receive a scholarship and recognition for one’s work”, he says.
More research needed
Dylan Mäenpää got interested in machine learning when he discovered that there is relatively little research about “federated learning”. This is a machine learning method used in organisations where sharing data is both sensitive and difficult – for example, between hospitals.
“Machine learning requires a lot of data in order to train models properly. It can often be hard to get hold of data, due to GDPR or the fact that you can’t just share data however you want to. So it’s often hard to train models well.”
Federated learning involves individual units – such as, in the context of healthcare, hospitals – training their models separately with local data. When they’re finished, the hospitals then share their models with each other, instead of sharing the data itself. Federated learning usually involves sharing the models by gathering them centrally. In his degree project, Dylan Mäenpää wanted to compare this method with another one where hospitals could communicate directly instead – via something known as a peer-to-peer model.
Dylan Mäenpää simulated different scenarios in which 100 different “places” learned to categorise numbers that they then shared with a central unit. He then compared the results with the peer-to-peer model, where the “places” shared their categorisation models with each other.
Lots of advantages
There are lots of advantages with the peer-to-peer system, according to Dylan Mäenpää.
“The results of these simulations show that federated learning could, theoretically speaking, be faster with this method. But when you just share the models in a central place, then several satellite units have to communicate with a central one. This makes the learning process slower. If the units simply communicate with each other, then you eliminate potential bottlenecks. That can make training the models quicker.”
Another advantage with this is reduced vulnerability. Using a central unit risks disrupting the teaching process if, for example, a technical problem occurs.
“There’s no ‘single point of failure’, as we say.”
Today, Dylan Mäenpää lives in Borås and works as an IT consultant at an e-commerce company in Gothenburg.
The title of Dylan Mäenpää’s degree project is “Towards Peer-to-Peer Federated Learning: Algorithms and Comparisons to Centralized Federated Learning”. It was presented on 1 June 2021.
More about the prize
Dylan Mäenpää’s degree project was awarded the prize due to its having ”shown in depth an understanding of existing methods, as well as creativity in designing new methods and evaluating them rigorously. Developing methods for distributed machine learning, where no single node in the system has access to all the sensitive data, has great potential for making machine learning and privacy compatible.”
The Christer Gilén scholarship has been awarded every year since 2019, and the prize money ranges from SEK 5000 to 20,000. The founder of the scholarship is Christer Gilén, who has donated SEK 100,000 to Linköping University’s Jubilee Foundation. The scholarship aims to encourage degree projects which produce useful results for industry or the public sector.
The scholarship is awarded in two areas: “statistics and machine learning” and “economic leadership, organisation och innovation”.