Grants awarded call Data Science 2018

A committee consisting of experts from Digital Humanities, Data Science & Systems Complexity (DSSC), the eScience Center of NWO and representatives of the CIT has honored seven projects within the framework of the call for proposals for data science 2018.

The committee received a total of 37 applications, a tripling in comparison with last year. The researchers of the awarded projects receive support from the data scientists of the CIT up to a maximum of 450 hours. The call was funded by the RUG's multi-annual plan 2016-2020. The next round of the call for proposals for data science starts in the second quarter of 2019.

Overview of awards

Eye in the sky: using satellite data to monitor real estate construction activity in global cities - Faculty of Spatial Sciences

This project aims to monitor the rate of real estate construction activity in 100 large cities across the world, on a monthly basis. The empirics of this large-scale analysis rely on object recognition in publicly available satellite imagery. The output helps to solve an international gap in data availability, which is crucial for better understanding value creating processes in the largest economic sector in terms of capital value -the real estate sector- and the timelier forecasting of economic growth or decline in cities and countries. Unlike existing data, the project’s output is set out to be public, high-frequency, and global. These qualities of the project support potential ground-breaking academic research, which may in turn solve key data availability problems that policy makers as well as public or private investors have. Promotion of results, and efforts in securing funding for future continuation of the project, can build on existing collaborations with top tier international real estate companies and policy research institutions. In addition, the PI will apply for additional NWO grants. Joint efforts with the CIT may include writing scientific papers, presenting project progress, and publishing standardized monthly reports on the rate of construction activity in cities for a global audience.

Track finding in matter-antimatter collisions using deep learning - KVI - Center for Advanced Radiation Technology

The objective behind this proposal is to provide a proof-of-principle to deploy Deep Neural Network (DNN) machine learning techniques to extract momenta of charged particles that are produced at extreme rates in matter-antimatter collisions at particle-accelerator facilities. The next-generation experiments require in-situ data processing at unprecedented interaction rates. We focus on evaluating track-recognition algorithms for the near-future antiProton ANnihilation at DArmstadt (PANDA) experiment that will be deployed at the Facility for Antiproton and Ion Research (FAIR) near Frankfurt, Germany. PANDA aims to search for new forms of bound states of quarks and study their interactions, thereby, providing unique data to give insight in the origin of colour confinement and the generation of the mass of visible matter. Labelled Monte Carlo (MC) data of the tracking detectors of PANDA will be used to demonstrate the feasibility and limitations of the proposed approach. Moreover, data from a running collider experiment, BESIII in China, will be used to benchmark the performance of the algorithm at low rates and low track multiplicities. The data will be provided by KVI-CART and the DNN will developed by data scientists at CIT. This project will serve as a springboard for a programme within a large-scale international collaboration.

Optimization of the yields of important bio-based chemicals with statistical approaches using large experimental datasets - Faculty of Science and Engineering

The Green Chemical Reaction Engineering (GCRE) group performs research on the development of highly intensified catalytic technology for biomass conversion to biofuels and bio-based chemicals. One specific area is the catalytic pyrolysis of biomass & waste to benzene, toluene and xylenes (BTX), on which GCRE has been co-operating with the company BioBTX BV in various projects since 2012. The performed experiments generated a vast amount of data, including the effect of process conditions on the BTX yield (first product group) and bio-oil yield (second product group). So far, trends in the data have been analysed using traditional methods. The unusual vast amount of data, however, requires the application of new and innovative approaches to obtain relations between BTX yield and process parameters. This research focusses on two specific areas:

1. Better quantification and trend analysis of the desired product group (BTX)
2. Develop a method for quantification of the second product group (bio-oil)

The quantification and trend analysis for BTX requires mostly statistical methods. The development of a method for quantification of the bio-oil combines statistical methods with a data interpretation method that is tailored towards specific components (polyaromatics; see Total Mass Sum in section 5a). This research is unprecedented and ground-breaking due to the unique combination of chemistry and data science. The desired outcome of this project would be i) a publication in a peer-reviewed journal and presentations at conferences, ii) identification of optimized process conditions to obtain highest BTX yields and to give directions to future research in the field of catalytic pyrolysis, iii) an improved and economic attractive route towards aromatics from natural resources, and iv) further growth of the BioBTX/RUG collaboration.

Data-driven infection management at the intensive care unit - improving quality of care through mining electronic health records - Faculty of Medical Sciences

Infections pose a major threat to critically ill patients on intensive care units (ICU) or can be the primary reason for ICU admission. Multidisciplinary infection management through antimicrobial stewardship (AMS) teams could optimize appropriate use of antimicrobials and diagnostics. However, traditional approaches are usually single event triggered (e.g. positive diagnostic results, specific antimicrobial prescriptions, complications, etc.), and disregard the complexity of ICU patients and clinical trends. Interventions still largely depend on descriptive statistics alone.

Advanced data science (e.g. machine learning) can bridge this gap using routinely collected data leading to data-driven stewardship in infection management. Visualization of large amounts of complex data and modeling can further facilitate communication and building up trust in pioneering applied data science approaches.

We have started collecting data from all adult patients at the ICU of the University Medical Center Groningen. We aim at using data science to optimize infection management (the right test and treatment for the right patient at the right time) and thereby to improve quality of care and patient outcome.

Guided learning of multi-task CNNs for joint place recognition and semantic segmentation - Faculty of Science and Engineering

Recently, Convolutional Neural Networks (CNNs) became the standard de facto for many tasks in Computer Vision. CNNs are neural networks with, usually, many hidden layers and a large number of parameters to be trained using large amount of labeled training data. They have been demonstrated successful in specific tasks such as object detection and classification, camera pose estimation, scene recognition, semantic segmentation and so on.

An important task in computer vision is visual place recognition, which is used in robotics to facilitate robot localization and loop-closure detection in simultaneous localization and mapping (SLAM) applications. Existing methods for place recognition based on CNNs exploit local visual cues extracted from pairs or triplet of images, while making limited use of contextual information about the objects in the scene. The aim of this proposal is two-fold. On one hand, we aim at exploring the capabilities of multiple CNNs to jointly learn effective representations for place recognition combining visual cues with information extracted from semantic segmentation. On the other hand, we are interested in investigating the possibility of training multi-task networks for place recognition and semantic segmentation, using the decisions of each network to improve the training of the other networks.

Intimate histories: a web-archaeological analysis of YouTube’s early history - Faculty of Arts

In this project, we will explore the possibilities of doing web-archaeological research in huge online platforms such as YouTube by finding new strategies of historical search in larger un- catalogued historical video data sets, and by performing a visual analysis of this data set through automated computer vision methods. By doing a case study, it will specifically trace the early history of YouTube as a site for domestic experiments in performing a high level of intimacy by everyday users experimenting with consumer media technologies such as the webcam and other mobile media devices. Through the data analysis of a relatively small corpus (4000 videos) we will be able to test the hypothesis that at the time there emerged a new genre, most likely recorded in indoor, domestic spaces, that preferred sharing intimate personal histories with a potential large global audience.

Automatic detection of key events in the budding yeast cell cycle based on Microscopy - Faculty of Science and Engineering

The study of dynamic processes inside living cells via time-lapse microscopy requires the capability to locate and track individual cells in consecutive images, as well as the annotation of key events of their cell division cycles. The unicellular microorganism Saccharomyces cerevisiae (budding yeast) is a widely used model system for the study of these processes, which are crucial for cancer development and ageing. Despite a plethora of software tools for segmenting and tracking single yeast cells in time-lapse imaging sequences, existing tools cannot detect when a new cell emerges on the surface of a mother cells (budding) or when a bud detaches from the mother cell (division). This implies that bud and division annotation are still carried out manually over huge numbers of cells, a very labor-intensive and time-consuming process. The goal of this project is to develop an image-processing software tool for the automatic annotation of budding and division events based on time-lapse microscopy of budding yeast. The resulting tool should make use of images obtained with simple microscopy techniques, which will make it widely useable. With time-lapse single-cell microscopy becoming the method of choice for studying the dynamics of cellular responses, the developed tool will be a major contribution to the yeast research community.