HPC-Lab Seminar at IDI, NTNU

This NTNU/IDI HPC-Lab seminar series is organized by Dr. Anne C. Elster. Inf the fall semster the seminar may be integrated with the course TDT 24: Numerical Methods and Parallel Environments, a master-level special-topics course run by Dr. Elster. Conteact her for details.

HPC-Lab Seminar at IDI, NTNU -- Spring 2012

"Image Organization and Visualization and the Self-Sorting Map" by Grant Strong, Memorial Univ., Canada

When/where: Friday Feb. 3, 15:15 (3:15pm) in ITV 454 (IDI's lunch room)

Abstract

Much of the work I have been a part of pertains to organizing collections of images based on their conceptual and visual similarities with the goal of facilitating a unique and intuitive view of those collections for the purpose of search. While there is no right or wrong answer for how images should be organized on a screen based on similarity, there are certainly better and worse ones. Through collaboration we have developed a technique that can generate an interactive search canvas of the results coming from an image search query. A set of images is pulled together from various concepts expanded out of the query using Wikipedia. In the back end, the organizing is performed by a Self-Organizing Map (SOM) that uses the feature vectors of the images (part concept, part content) to compute screen positions. Since SOM training is a bottleneck, we do it in parallel in the GPU. I will give an overview of this process and demo some results.

My newest work is on an alternative to the SOM for this type of organizational problem. Termed the Self-Sorting Map, this algorithm is an attempt to eliminate the overheads associated with the SOM by simply shuffling items in cells of a grid based on their relative similarities. The desired output is something roughly akin to "2D sorting", in which items are placed in the grid in such a way that proximity reflects similarity. We are hoping to achieve results similar to the SOM that are quicker to obtain and simpler use. The algorithm has been devised to be parallel to take advantage of modern hardware. I will discuss the algorithm and show the preliminary results with the test problems we have been using to evaluate it.

About the speaker

Grant Strong, a senior PhD student from the CS Dept. at Memorial University in Canada (http://www.mun.ca/computerscience/) is visiting IDI's HPC-Lab for 3 this spring semester.
"Computational Physics with X-rays: Ptychography and X-ray Diffraction Visualization" by Thomas L. Flach

Thomas Falch will be talking about his fall project + master's thesis.

Time and place: 14:15-15:00 in 454 (Wed Feb 22)

Abstract:

Ptychography is a promising new X-ray microscopy technique. The reconstruction the real space image from the output of the microscope does, however, require massive computational effort. At the same time, more and more computational power is made available. Recently, however, this has been primarily in the form of parallel computers. Writing multithreaded programs taking full advantage of these architectures is difficult, even for trained programmers. In this project we will look at a program performing ptychographical reconstruction. We will start with a Python version, translate it to C, and adapt it to achieve high performance on a modern multi-core processor. To do this we parallelize it using thread-pools and SIMD instructions, and apply various optimization techniques, like array merging, expression rewriting and loop unrolling. We are able to achieve a speedup of 18.01 over the initial Python version, and 6.50 over the initial, serial C version, on a 4 core processor with simultaneous multithreading. This is a significant result, the speedup makes it possible to use the result of one X-ray microscopy experiment help guide the next one without prohibitive waiting times, essentially turning the analysis of the X-ray microscopy results into an interactive rather than offline job. With the introduction of fast area detectors, X-ray diffraction experiments now routinely yield data sets on the order of several gigabytes when studying nanostructured materials with the method known as "reciprocal space mapping". Due to complicated scattering geometries, the data points do not fall on any grid, nor is the density of points constant in all directions. These facts, along with the large data volumes, makes creating meaningful visualizations challenging. The overall goal of this project is to develop methods to create high-quality visualizations of three-dimensional X-ray diffraction data. This will be done by extending and modifying volume rendering techniques such as ray casting. Due to the computationally intensive nature of the problem, modern multi-core processors and programmable GPUs will be considered.

About Thomas Falch:

Thomas is a Master of Technology student at the IDI HPC-Lab with Dr. Anne C. Elster as his main advisor and Dr. Dag W. Breiby as his co-advisor. He joined NTNU as a nano-tech student, but decided he liked doing computer science, so he then joined IDI. His project is a collaboration with Dr. Dag Breiby's research group in condensed matter physics at NTNU. Thomas has expressed interest in a PhD at HPC-Lab.

Announcements:

Next time: Dr. Bjørn Angelsen will give an invited talk about HPC issues in Medical Technology Finally: Those who missed HPC-Lab visitor Grant Strong's talk last time, feel free to stop by his office at 417 and discuss anything related to his "Image Organization and Visualization and the Self-Sorting Map" talk. Grant will be here for another 2 months.
Bjørn Angelsen will give this week's HPC-Seminar on "Beregningsutforninger i medisinsk ultralydavbildning" loosely translated: "Computational Challenges in Medical Ultrasound Imaging"

Where and when: ITV 454 @ 13:15-14 Friday April 27, 2012


Exernal talks/Informal gatherings - 2012



This page is maintained by : HPC-Lab leader Dr. Anne C. Elster, elster-at-idi.ntnu.no. All photos are done and copyrighted by her. Please contact her if you wish to use them.

It was last updated on Aug 30, 2010. Comments welcome.


~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~