5.1 million euros made available for leading research in astronomy, computer science and mathematics
Enabling zero-energy wireless communication with ambient light (compartment 1): Dr M.A. (Marco) Zuñiga Zamalloa.
‘Humans already use 50% more energy moving information than moving airplanes around the world. Communication is central to our societies but it is taking a toll on the earth. We want to use a free, abundant and natural resource for wireless communication: sunlight. Similar to the way you can use a mirror to communicate by reflecting light, we will enable objects to change their reflection properties to send information, but without you noticing it! In this manner, buildings, cars and other objects in our cities will be able to talk to each other using daylight, an eco-friendly solution.’
Pixel-free Deep learning (compartment 2): Dr J. C. (Jan) van Gemert.
‘The recent breakthrough in deep learning is rivaling human image understanding by using massive data sets and fast hardware to learn image filters represented by pixel weights. Recent research investigates filters that are independent to image rotations, camera viewpoint changes, or variations in object size. To date, however, little work has addressed how to make deep networks independent of the number of pixels in an image. As a consequence, when the image resolution changes then the complete deep network architecture also has to change, which is a difficult and time consuming trial-and-error process. My project, "Pixel-free Deep learning", remedies this gap by injecting CNNs with scale-space theory where resolution no longer depends on the number of pixels and frees the network from using pixel weights. Traditional models cannot learn filter resolution because the number of pixels is a discrete measure that cannot be optimized through gradient descent. In contrast, scale-space theory models image resolution with a single continuous parameter, which can readily be learned by a deep network. The benefits are a more flexible network design, a strong theoretical foundation for visual deep learning, leading to robust and better understandable models. I will investigate (1) resolution learning for image classification (2) resolution learning for per-pixel predictions; (3) theoretical properties and understanding; (4) computational efficiency. This proposal challenges the current scientific foundations of deep convolutional networks where a strong simplification of manual network design is a game-changer for industrial applications of deep learning.’
Interacting Spreading Processes on Interdependent Social Networks (compartment 2): Dr H. (Huijuan) Wang.
‘There’s an explosion of online social networks, each of which supports the spread of information, opinions and behaviours, the so-called social contagion. A user’s activity in one network may influence not only activities of his/her friends within that network but also his/her activities in other networks. Lacking understanding of such interacting patterns between networks, we are unable to explain nor to control the emergence of collective behaviours such as cascading failures, social riots and the popularization of (mis)information, innovation and policy adoption.’ Huijuan Wang’s overarching goal is to discover and model the unknown mechanisms of the spread or contagion of user activities on interdependent networks and provide foundational understanding of their emergent effect. This project will push the boundary of network science, which focuses on single theoretical spreading processes, by addressing the interaction between spreading processes that explains not only the average but also the diversity of user activities i.e. the emergence of active users or networks, benchmarked by the data.