Mari W. Buche (DataS/SBE) attended the HERS Institute July 8-20, at Bryn Mawr College, Pennsylvania, Along with 63 women from a variety of colleges and universities across the country, Buche participated in professional development sessions focused on fine-tuning leadership skills unique to higher education. Visit hersnetwork.org for more details.
Researcher: Jianhui Yue, PI, Assistant Professor, Computer Science
Sponsor: National Science Foundation, SHF: Small: Collaborative Research
Amount of Support: $192, 716
Duration of Support: 3 years
Abstract: Emerging nonvolatile memory (NVM) technologies, such as PCM, STT-RAM, and memristors, provide not only byte-addressability, low-latency reads and writes comparable to DRAM, but also persistent writes and potentially large storage capacity like an SSD. These advantages make NVM likely to be next-generation fast persistent storage for massive data, referred to as in-memory storage. Yet, NVM-based storage has two challenges: (1) Memory cells have limited write endurance (i.e., the total number of program/erase cycles per cell); (2) NVM has to remain in a consistent state in the event of a system crash or power loss. The goal of this project is to develop an efficient in-memory storage framework that addresses these two challenges. This project will take a holistic approach, spanning from low-level architecture design to high-level OS management, to optimize the reliability, performance, and manageability of in-memory storage. The technical approach will involve understanding the implication and impact of the write endurance issue when cutting-edge NVM is adopted into storage systems. The improved understanding will motivate and aid the design of cost-effective methods to improve the life-time of in-memory storage and to achieve efficient and reliable consistence maintenance.
Publications:
Pai Chen, Jianhui Yue, Xiaofei Liao, Hai Jin. “Optimizing DRAM Cache by a Trade-off between Hit Rate and Hit Latency,” IEEE Transactions on Emerging Topics in Computing, 2018. doi:10.1109/TETC.2018.2800721
Chenlei Tang, Jiguang Wan, Yifeng Zhu, Zhiyuan Liu, Peng Xu, Fei Wu and Changsheng Xie. “RAFS: A RAID-Aware File System to Reduce Parity Update Overhead for SSD RAID,” Design Automation Test In Europe Conference (DATE) 2019, 2019.
Pai Chen, Jianhui Yue, Xiaofei Liao, Hai Jin. “Trade-off between Hit Rate and Hit Latency for Optimizing DRAM Cache,” IEEE Transactions on Emerging Topics in Computing, 2018.
By Karen S. Johnson
With close to 2,000 working satellites currently orbiting the Earth, and about a third of them engaged in observing and imaging o
ur planet,* the sheer volume of remote sensing imagery being collected and transmitted to the surface is astounding. Add to this images collected by drones, and the estimation grows quite possibly beyond the imagination.
How on earth are science and industry making sense of it all? All of this remote sensing imagery needs to be converted into tangible information so it can be utilized by government and industry to respond to disasters and address other questions of global importance.
In the old days, say around the 1970s, a simpler pixel-by-pixel approach was used to decipher satellite imagery data; a single pixel in those low resolution images contained just one or two buildings. Since then, increasingly higher resolution has become the norm and a single building may now occupy several pixels in an image.
A new approach was needed. Enter GEOBIA– Geographic Object-Based Image Analysis— a processing framework of machine-learning computer algorithms that automate much of the process of translating all that data into a map useful for, say, identifying damage to urban areas following an earthquake.
In use since the 1990s, GEOBIA is an object-based, machine-learning method that results in more accurate classification of remotely sensed images. The method’s algorithms group adjacent pixels that share similar, user-defined characteristics, such as color or shape, in a process called segmentation. It’s similar to what our eyes (and brains) do to make sense of what we’re seeing when we look at a large image or scene.
In turn, these segmented groups of pixels are investigated by additional algorithms that determine if the group of pixels is, say, a damaged building or an undamaged stretch of pavement, in a process known as classification.
The refinement of GEOBIA methods have engaged geoscientists, data scientists, geographic information systems (GIS) professionals and others for several decades. Among them are Michigan Tech doctoral candidate James Bialas, along with his faculty advisors, Thomas Oommen(GMERS/DataS) and Timothy Havens (ECE/DataS). The interdisciplinary team’s successful research to improve the speed and accuracy of GEOBIA’s classification phase is the topic of the article “Optimal segmentation of high spatial resolution images for the classification of buildings using random forests” recently published in the International Journal of Applied Earth Observation and Geoinformation.
The team’s research started with aerial imagery of Christchurch, New Zealand, following the 2011 earthquake there.
“The specific question we looked at was, how do we translate the information we get from the crowd into labels that are coherent for an object-based image analysis?” Bialas said, adding that they specifically looked at the classification of city center buildings, which typically makes up about fifty percent of an image of any city center area.
After independently hand-classifying three sets of the same image data with which to verify their results (see images below), Bialas and his team started looking at how the image segmentation size affects the accuracy of the results.
“At an extremely small segmentation level, you’ll see individual things on building roofs, like HVAC equipment and other small features, and these will each become a separate image segment,” Bialas explained, but as the image segmentation parameter expands, it begins to encompass whole buildings or even whole city blocks.
“The big finding of this research is that, completely independent of the labeled data sets we used, our classification results stayed consistent across the different image segmentation levels,” Bialas said. “And more importantly, within a fairly large range of segmentation values, there was pretty much no impact on results. In the past several decades a lot of work has done trying to figure out this optimum segmentation level of exactly how big to make the image objects.”
“This research is important because as the GEOBIA problem becomes bigger and bigger—there are companies that are looking to image the entire planet earth per day—a massive amount of data is being collected,” Bialas noted, and in the case of natural disasters where response time is critical, for example, “there may not be enough time to calculate the most perfect segmentation level, and you’ll just have to pick a segmentation level and hope it works.”
This research is part of a larger project that is investigating how crowdsourcing can improve the outcome of geographic object-based image analysis, and also how GEOBIA methods can be used to improve the crowdsourced classification of any project, not just earthquake damage, such as massive oil spills and airplane crashes.
One vital use of of crowdsourced remotely sensed imagery is creating maps for first responders and disaster relief organizations. This faster, more accurate GEOBIA processing method can result in more timely disaster relief.
*Union of Concerned Scientists (UCS) Satellite Database
Thomas Oommen (DataS/GMES/EPSSI) is Principal Investigator on a project that has received a $39,999 research and development grant from the US Department of State. The project is titled, “Developing and Improving Disaster Management Studies Course in India.” This is a one-year project.
Timothy Havens (CC/ICC) was General Co-Chair of the 2019 IEEE International Conference on Fuzzy Systems in New Orleans, LA, June 23 to 26. At the conference, Havens presented his paper, “Machine Learning of Choquet Integral Regression with Respect to a Bounded Capacity (or Non-monotonic Fuzzy Measure),” and served on the panel, “Publishing in IEEE Transactions on Fuzzy Systems.”
Three additional papers authored by Havens were published in the conference’s proceedings: “Transfer Learning for the Choquet Integral,” “The Choquet Integral Neuron, Its PyTorch Implementation and Application to Decision Fusion,” and “Measuring Similarity Between Discontinuous Intervals – Challenges and Solutions.”
ICC Director Tim Havens (DataS) presented an invited talk, “Explainable Deep Fusion,” at the Technological University of Eindhoven, The Netherlands, on May 7, 2019.
Like a winning trivia team, sensor fusion systems seek to combine cooperative and complementary sources to achieve an optimal inference from pooled evidence. In his talk, Havens introduced data-, feature-, and decision-level fusions and discussed in detail two innovations he has made in his research: non-linear aggregation learning with Choquet integrals and their applications in deep learning and Explainable AI (XAI).
Tim Havens (CS/ICC) coauthored the article, “Enabling Explainable Fusion in Deep Learning with Fuzzy Integral Neural Networks,” which was accepted for publication in the journal IEEE Transactions on Fuzzy Systems.
Citation: M.A. Islam, D.T. Anderson, A. Pinar, T.C. Havens, G. Scott, and J.M. Keller. Enabling explainable fusion in deep learning with fuzzy integral neural networks. Accepted, IEEE Trans. Fuzzy Systems.
Abstract: Information fusion is an essential part of numerous engineering systems and biological functions, e.g., human cognition. Fusion occurs at many levels, ranging from the low-level combination of signals to the high-level aggregation of heterogeneous decision-making processes. While the last decade has witnessed an explosion of research in deep learning, fusion in neural networks has not observed the same revolution. Specifically, most neural fusion approaches are ad hoc, are not understood, are distributed versus localized, and/or explainability is low (if present at all). Herein, we prove that the fuzzy Choquet integral (ChI), a powerful nonlinear aggregation function, can be represented as a multi-layer network, referred to hereafter as ChIMP. We also put forth an improved ChIMP (iChIMP) that leads to a stochastic gradient descent-based optimization in light of the exponential number of ChI inequality constraints. An additional benefit of ChIMP/iChIMP is that it enables eXplainable AI (XAI). Synthetic validation experiments are provided and iChIMP is applied to the fusion of a set of heterogeneous architecture deep models in remote sensing. We show an improvement in model accuracy and our previously established XAI indices shed light on the quality of our data, model, and its decisions.
The Office of the Provost and Senior Vice President for Academic Affairs has announced that Dr. Timothy Schulz (DataS), professor of Electrical and Computer Engineering, has been named a 2019 University Professor.
The University Professor title recognizes faculty members who have made outstanding scholarly contributions to the University and their discipline over a substantial period of time. University Professors will not exceed 2% of the total number of tenured and tenure-track faculty at Michigan Tech. This year, two professors were awarded the title of University Professor. The second recipient is Dr. Kathleen Halvorsen, professor of Natural Resource Policy in the Department of Social Sciences.
The confidential process for selecting recipients spans the academic year and recipients for each award are notified in mid-May. Additional details regarding the award and selection procedures can be found on the provost’s website: mtu.edu/provost/faculty/awards.
by Michael R. Meyer, Director William G. Jackson CTL
College of Engineering Dean Janet Callahan has selected Tim Schulz (ECE) as the final member of the 2019 Deans’ Teaching Showcase. As a teacher he is widely acknowledged as one of the ECE departments best, with his friendly, humorous style and his devotion to his students’ learning. But Schulz’s selection here is, according to Associate Dean Leonard Bohmann for his “leadership in using technology to deliver technical material in electrical and computer engineering.”
Starting in 2012, Schulz created a series of 10 to 15 minute videos collectively titled “Electric Circuits” and posted them on YouTube. Though he created them with his EE2111 (Electric Circuits 1) class in mind, they are reaching a much wider audience. In fact, one titled “Introduction to Thevenin Equivalent Circuits” has gotten more than 152,000 views.
Since that time, Schulz has also developed a phone app of randomized electric circuit problems to use in this course. He develops these aids so students can develop a mastery of the course material. As one student noted, “The videos and the infinite practice problems were the most helpful. As much as I hate to say this, the quizzes were also helpful.”
In his courses, Schulz develops from scratch his own interactive web-based approach to homework sets and quizzes, taking full advantage of the capabilities of Canvas and writing his own scripts for generating homework problems with randomized parameters. His colleagues recognize this, and some have adopted Schulz’s materials when they teach the same classes.
Most recently, Schulz has taken the lead in developing new courses for the online MSEE program with a focus on communications and signal processing, in partnership with Keypath Education, Inc. He developed and is teaching for the second time, EE5300, Mathematical and Computational Methods in Engineering, which is the entry point into the program.
His course engages students through a series of interactive MATLAB computational exercises which meet modern standards for online course delivery and are breaking new ground for the ECE Department.
Students find this approach to be very helpful. One said, “The canvas structure paired with the lecture truly was a great combination. The prep work must have been substantial but was well worth it.”
Another provides even broader praise of both Schulz and the course by saying, “The course is excellent and engaging. Overall, I think this class is a must for any student wishing to have a solid starting foundation in graduate studies in engineering. Dr. Schulz is an outstanding professor with extensive research and professional experience and I would totally recommend students to take this class.”
Schulz is currently developing the third course for the online MSEE program, EE5500 Probability and Stochastic Processes, which will be taught for the first time this summer. He agrees that developing an online course is much more rigorous then teaching face-to-face, saying “You need to do more planning of how to approach a topic. You don’t have the ease of correcting an approach (or even an equation) in real time, so it is a much more deliberate process.”
However, this higher level of rigor is a challenge he enjoys; he’s already signed on to develop his next course, EE5521 Detection and Estimation Theory, which will be offered online for the first time sometime in 2020-2021 academic year.
Callahan emphasizes that it’s really about the technology enabling better learning. In her words, “Tim Schulz’s effective use of technology shows that student learning and satisfaction can both increase with the use of modern tools.”
Schulz will be recognized at an end-of-term luncheon with other showcase members and is now elgible for one of three new teaching awards to be given by the William G. Jackson Center for Teaching and Learning this summer recognizing introductory or large class teaching, innovative or outside the classroom teaching methods, or work in curriculum and assessment.
Timothy Havens (ECE) is the principal investigator on a research and development project that has received $96,643 from the Naval Surface Warfare Center. Andrew Barnard (ME-EM) is the Co-PI on the project, which is titled, “Localization, Tracking, and Classification of On-Ice and Underwater Noise Sources Using Machine Learning.”
This is the first year of a potential three-year project totaling $299,533.
Tech Today, March 7, 2019