Dr Stefan Vlaski tells us about his research at the intersection of machine learning, optimisation, and network science.

Technologies based on machine learning and AI have an increasing impact on our daily lives, and there is huge interest and research activity in this area, as well as great potential for collaboration within and across disciplines. Stefan’s expertise will complement and enhance our research in this fast-moving field.

It is becoming more challenging, yet more important to understand the systems we involve in our lives, make them transparent, and have the ability to trust them. That is the ultimate objective of my research, and the potential for far-reaching positive impact is really exciting.” Stefan Vlaski

Stefan completed his undergraduate degree in Germany, before moving to the US for his postgraduate studies at UCLA.

“I have always been into mathematics, but I wanted to apply it to make a real-world impact, and that’s what led me to study engineering. Throughout my undergraduate degree, I was exposed to signal processing — the science of extracting information from signals and data. I was fascinated by how it can be applied to all aspects of our life, from communication systems, to speech and image recognition algorithms in many modern phones and cars, or to time-series data describing the growth of financial stocks or even pandemics.”

“My current research centres around the development of algorithms for decentralised optimisation and machine learning. Optimisation is all about designing the “best” solutions to problems. This could be maximising the transmission rate in a communications system, or allocating power across a smart grid, or managing a financial investment portfolio.

In the context of machine learning, the objective is most commonly to find a model (such as a neural network), which describes the data best, and allows us to make predictions with the maximum accuracy.”

Why is decentralising machine learning important?

“In most cases, the resulting optimisation algorithms are centralised, meaning that both data and computational power resources need to be available at a central location. Fusing and processing data at a central location is generally neither efficient nor feasible — for example the fusion centre is a potential point of failure. If you tried using a Facebook service over the last couple of days you will have noticed this.

However, as technology becomes pervasive in all aspects of society and our lives, both computational power and data are increasingly available as dispersed locations such as our mobile devices, which collect data about us every day. Any single device in isolation may have limited data, and limited computational power, but added together this information reveals a lot about us.

The information has the potential to significantly enrich our lives. Our interests can generate product recommendations, our movements can predict traffic patterns, and our search results can predict pandemics. But for the very same reason, we may be uncomfortable sharing this private information with the device manufacturer.

Federated learning models can give great performance and have some advantages over fusing the raw data — where the data is sent directly to the fusion centre and all processing is done there — but they still rely on a central coordinator.

The purpose of my research is to develop decentralised algorithms for optimisation and learning, where the learning happens without any central coordination. Instead the network of “agents” (i.e., sensors, mobile devices) do the local processing, followed by peer-to-peer interactions.

When properly designed, this network system can solve the underlying optimisation or learning problem just as well as a powerful central fusion centre. And because they are decentralised, they tend to be more flexible in the presence of communication constraints they are more robust against node and link failures, and they offer higher levels of privacy protection.”

Future challenges

“Because there is no longer a central coordinator, the dynamics of these systems become much more interesting. It is not obvious how their performance will compare to federated architectures. They take on a mind of their own, and many unforeseen things can happen if we’re not careful. I’m interested in devising such systems optimally, understanding their behaviour, advantages and drawbacks. 

Applications of optimisation and machine learning, particularly in the context of multi-agent systems, continue to spill into all aspects of our lives. The content we are presented with on social networks, and on-demand platforms, is largely curated by machines. Autonomous cars and vehicles are full of artificial intelligence. Manufacturing is being changed by robotics. Medical experts are increasingly relying on machine learning to drive research and diagnostics. As the range of applications and stakes grow, so does the complexity of these systems. This means that it is becoming more challenging, yet more important to understand the systems we involve in our lives, make them transparent, and have the ability to trust them. That is the ultimate objective of my research, and the potential for far-reaching positive impact is really exciting.” 

A collaborative environment

In addition to his research, Stefan will be tutoring undergraduates this term, and he is also looking forward to supervising student projects in the area of machine learning, optimisation and networks. 

“One of the great things about Imperial is the density of cutting-edge research focused on the areas of science and engineering, and what excites me about optimisation and machine learning is the applicability to many areas of science and engineering, some of which we already know, and many of which are surely there to be discovered. So, I am very much looking forward to interacting with students and researchers from all across Imperial, and hopefully finding an exciting new application for my algorithms.”

Discovering London

“As a city to live in, London is unmatched. I’m still surprised every time I step from a loud and busy street into Hyde Park and can hear birds singing after a couple of steps. 

I enjoy sports, and normally this means running and playing tennis, but when I came to London I learned that people sail on the Thames, so I’ve spent a couple of weekends attempting to sail, capsizing, and swimming in the river.

From museums to restaurants, parks and different boroughs, there is so much to discover in this city, and I’m very much looking forward to doing that in my free time.”

Welcome to the EEE Community, Stefan!