In the future, different intelligent systems will need to share data and experiences with each other to become good enough for certain tasks. RISE Industrial PhD student Martin Isaksson's research is an important step on the way. This area is highly relevant for Ericsson and the development of its 5G but lessons learned along the way might open for many more solutions.
The conference room on the sixth floor of RISE’s Kista office is bursting at the seams. Even the supply of folding chairs is almost exhausted. Representatives from RISE and Ericsson are gathered in fevered expectation. They are here to listen to Martin Isaksson explain the concept behind the subject he has chosen for his doctoral thesis – decentralised machine Learning.
“I decided to begin at the end, with the final objective,” says Martin Isaksson. “Then, I hope that the details will make themselves known along the way and the common thread will appear. I have thrown myself into this with a mixture of delight and terror.”
Martin, who comes from Ericsson Research, had not intended to continue on to doctoral studies until the offer of funding turned up from WASP, the Wallenberg Artificial Intelligence, Autonomous Systems and Software Program.
“I truly appreciate this opportunity to be able to focus on the subject and really learn about it in depth. That would never have been possible without WASP’s support.”
Privacy aspects needs to be considered
Martin’s background is in Ericsson’s LTE Radio Access Network division. It was natural that he would bring the knowledge acquired there to bear on his PhD project, which deals with decentralised machine learning. According to Martin Isaksson, there are two main reasons why he chose this subject in particular.
“There is so much sensitive data on the edge of the network, for example in mobile telephones, data that can’t be incorporated in traditional machine learning where it is sent for central analysis. The amount of data in itself does not allow this, although the primary aspect is that of privacy; we cannot use private data such as texts or images for centralised learning. We therefore need to look at opportunities to handle learning locally on the edge of the network.”
The challenge is to get the net to learn in unison, despite the fact that the data can’t be sent.
“One idea is to send pre-trained models, or updates to a pre-trained model, instead of data,” reasons Martin Isaksson. “On the other hand, this brings its own challenges with regard to security and privacy, as well as statistical and system challenges.”
Two supervisors from RISE
Martin’s supervisors for the project are Seif Haridi, chief scientific advisor at RISE SICS and professor at KTH Royal Institute of Technology, and Daniel Gillblad, senior researcher into AI at RISE SICS. According to Martin Isaksson himself, having Seif and Daniel as supervisors is a decisive factor for the success of the project.
“I have been enormously lucky in being assigned two such experienced and competent supervisors.”
Daniel Gillblad believes that the subject chosen by Martin is central to making machine learning systems and AIs function on a large scale in society.
“Different machine learning systems will need to exchange what they have learned from the shared data, to learn from each other’s experiences, if they are to become good enough for certain tasks. Martin’s research is one step along the way to realising this ambition”
Many years of collaboration
The collaboration between RISE and Ericsson goes back as far as 1985, when Ericsson and the then Televerket took the initiative to establish the Swedish Institute for Computer Science (SICS) in order to meet the need for industrial research.
“The list of our current areas of collaboration is a very long,” confirms Sverker Janson, director of the Computer Systems Laboratory at RISE. “We collaborate on everything from AI, machine learning and scalable systems to security, Internet of Things and networks. Cooperation covers both strategic research and its application in products, including both concrete assignments and collaborative projects financed by Vinnova, the EU or The Swedish Foundation for Strategic Research (SSF).”
It was Rickard Cöster, an expert at Ericsson Research, who first convinced Martin Isaksson to apply to the research organisation. He too is convinced that the collaboration between Ericsson and RISE is beneficial to both organisations.
“As well as having long experience in our domain, RISE SICS is an important partner for Ericsson in the field of AI and machine learning. Martin’s field is highly relevant to Ericsson’s development of 5G, but even more so generally with regard to the distributed intelligent systems of the future; so, naturally there is a great deal of interest and support from both parties.”
Will contribute to Ericsson’s business
The project’s contribution to Ericsson’s business was obvious to Martin Isaksson from the beginning, when he realised that the challenges facing the LTE division could be addressed with the help of AI. It is however the specialisation in the subject that Martin is most looking forward to.
“For me, the lessons learned along the way are what’s important. I feel like I’m perched at the top of a rollercoaster and I’m looking forward to a thrilling ride.”