02001 Research Immersion Projects, 2018

Sparse Matrix-Free Matrix-Vector Products on GPUs

• Contact: Allan Peter Engisg-Karup (apek@dtu.dk)
• Description: Study how to optimally map a sparse matrix-vector product efficient to a Graphics Processing Unit (GPU) to achieve massively parallel performance on such devices. The algorithm is quite simple and performs the operation A*x=b, where the b vector is to be computed as the action of a sparse matrix that is not assembled through product with an input vector x. The project is relevant for many scientific computing implementations and is an often used operation in many codes, however, the approach deviates from implementations in de factor standard libraries in that the matrix A is never assembled. Thus, this approach is matrix-free as it is based on no assembly of a matrix, and instead the unknown of the sparse matrix is stored in a way that can be exploited efficiently utilising the memory hierarchy on the GPU. The project is therefore an excellent opportunity to learn about high-performance computing and can be connected to an practical state-of-the-art large-scale hydrodynamics applications code that is already massively parallel and massively scalable. This code has been developed at DTU Compute (http://www.compute.dtu.dk/~apek/oceanwave3d) and used for both research and industrial applications in the last decade.

Security in the era of IoT

• Contact: Michele de Donno (mido@dtu.dk)
• Description: The Internet of Things (IoT) revolution has not only carried the promise of interconnecting a whole generation of "dumb" devices, but also brought to the Internet billions of badly protected and easily hackable objects. Indeed, this sudden flooding of insecure devices brought back to the top older threats, such as Distributed Denial of Service (DDoS) attacks. Our research group at DTU Compute is working on this issue, both analyzing the current state of the IoT Security (further details can be found: https://www.hindawi.com/journals/scn/2018/7178164/abs/) and trying to develop a possible solution to solve, or at least mitigate, the problem (further details can be found: https://link.springer.com/chapter/10.1007/978-3-319-70578-1_7).
• Based on background and skills of the student(s), different projects are available in this research area, ranging from theoretical to practical ones.

Visualization in Mobile Apps for Behavioural Activation

• Contact: Darius Adam Rohani (daroh@dtu.dk)
• Description: Do you want to take part in a PhD project and earn 5 ECTS? Do you find programming interactive visualizations in mobile apps interesting? Then this is the perfect research course for you! Your part of the project will consist of creating an interactive visualization for an app (MUBS). The app is a personal BA recommender-system for people with depression. BA is a therapy method that consists of planning activities and getting them done. More information about the overall aim of the PhD projectcan be found at: http://www.cachet.dk/research/research-projects/radmis The app is being programmed in Xamarin Forms (Programming languages: C# and xaml).
• Schedule:
  • Week 1: Learn how to use Xamarin SkiaSharp
  • Week 2 + 3: Prototype and implement visualizations with BA focus
  • Week 4: Do a small qualitative study and write a report of the app and the findings.
• Requirements: C# programming.

Building Digital Models of Physical Objects

• Contact: Andreas Bærentzen (janba@dtu.dk)
• Description: You will create digital models from a physical objects of your own choosing. The Section for Image Analsys and Computer Graphics has developed a structured light scanner (SeeMaLab Scanner). The aim of the project is to scan an object in order to reconstruct a 3D model using our in-house software. However, the 3D model is just the starting point. Your next task is to photograph the object and create a high fidelity, lighting independent texture from the photographs. The end result should be a digital model that could be used as a stand in for the chosen object in computer graphics applications such as VR. Depending on the interest of the students, this project could be more applied using software such as Blender or software could be developed specifically for the texture map generation. In either case, the goal is to develop a well documented pipeline.

Deep learning for multi-spectral CT reconstruction

• Contact: Wail Mustafa (wamus@dtu.dk), Christian Kehl (chke@dtu.dk), Søren Gregersen (sorgre@dtu.dk)
• Description: In Computed Tomography (CT), a single 2D image is reconstructed from a series of 1D x-ray projections, sampled at different rotational angles. Deep learning can be used to improve the image quality when the projection data are under-sampled (i.e. data is obtained from only a few rotational angles). In a previous work, a variant of the U-net architecture was used for the image reconstruction. However, the architecture assumes a single energy mode and does not utilize the energy spectrum provided by modern x-ray detectors. The aim of this project is to investigate the incorporation of the spectrum in the network architecture. The student is expected to review the state of the art architectures/approaches for multi-channel data first. The new architecture should then be implemented and evaluated. The existing architecture for single- channel CT data, implemented in Keras, can be used a starting point. Moreover, part of the project is to generate sufficient amount of spectral data for training and testing. The student will be provided the tools to generate the data.