DISSERTATION DEFENSE

       Department of Computer Science and Engineering

University of South Carolina 

Author : Marios Xanthidis

Advisor : Dr. Ioannis Rekleitis and Dr. O'Kane

Date : Feb 16, 2022

Time 3:00 pm

Place : Seminar room 2277, Storey Innovation Building

Abstract

The holy grail of robotics is producing robotic systems capable of efficiently executing all the tasks that are hard, or even impossible, for humans. Humans, undoubtedly, from both a hardware and software perspective, are extremely complex systems capable of executing many complicated tasks. Thus, the complexity of state-of-the-art robotic systems is also expected to progressively increase, with the goal to match or even surpass human abilities. Recent developments have emphasized mostly hardware, providing highly complex robots with exceptional capabilities. On the other hand, they have illustrated that one important bottleneck of realizing such systems as a common reality is real-time motion planning.

This thesis aims to assist the development of complex robotic systems from a computational perspective. The primary focus is developing novel methodologies to address real-time motion planning that enables the robots to accomplish the goals safely and provide the building blocks for developing robust advanced robot behavior in the future. The proposed methods utilize and enhance state-of-the-art approaches to overcome three different types of complexity:

1. Motion planning for high-dimensional systems. RRT⁺, a new family of general sampling-based planners, was introduced to accelerate solving the motion planning problem for robotic systems with many degrees of freedom by iteratively searching in lower-dimensional subspaces of increasing dimension. RRT⁺ variants computed solutions in real-time, orders of magnitude faster compared to state-of-the-art. Experiments in simulation of kinematic chains up to 50 degrees of freedom, and the Baxter humanoid robot validate the effectiveness of the proposed technique.

2. Underwater navigation for robots in cluttered environments. AquaNav, a real-time navigation pipeline for robots moving efficiently in challenging, unknown, and unstructured environments, was developed for Aqua2, a hexapod swimming robot with complex, yet to be fully discovered, dynamics. AquaNav was tested offline in known maps, and online in unknown maps utilizing vision-based SLAM. Rigorous testing in simulation, in-pool, and open-water trials show the robustness of the method on providing efficient and safe performance, enabling the robot to navigate by avoiding static and dynamic obstacles in open-water settings with turbidity and surge.

3. Active perception of areas of interest during underwater operation. AquaVis, an extension of AquaNav, is a real-time navigation technique enabling robots, with arbitrary multi-sensor configurations, reach safely their target, while at the same time observing from a desired proximity multiple areas of interest. Extensive simulations show a safe behavior, and strong potential for improving underwater state estimation, monitoring, tracking, inspection, and mapping of objects of interest in the underwater domain, such as coral reefs, shipwrecks, and marine life.

Abstract

The talk focuses on insightful problem solving and on scientific discovery as the most sophisticated form of insight. I will review several classical insight problems and will offer a conjecture that symmetry of the representation is what is common to all of them. Next, I will describe two phenomena from visual perception, 3D shape reconstruction and figure-ground organization, which also critically depend on symmetry of visual representation. This way, 3D vision can be considered the most elementary, but at the same time, ubiquitous form of insightful problem solving. In the third part of the talk, I will discuss the fundamental role symmetry plays in mathematics and physics including formulation of the Natural Laws. If time allows, I will conclude by describing how humans solve combinatorial optimization problems .

Bio

Professor Pizlo received his PhD in electrical and computer engineering from the Institute of Electron Technology, in Poland in 1982, and another PhD in psychology from the University of Maryland, College Park in 1991. He was a Professor of Psychology at Purdue University from 1991 to 2017 and is now a Professor of Cognitive Sciences at UC Irvine. He published 3 books on visual perception of shapes and space and his new book on problem solving will come out this Summer. His research on vision combines projective geometry and symmetry with inverse problems and regularization methods to solve them. His work on problem solving focuses on combinatorial optimization problems. 

Location:

In person

Swearingen Engineering Center in Room 2A31

Virtual MS Teams

https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTY5ODJjOTgtOTZhYi00OTJmLTljYTgtNjlkYzMxZjI5NjVk%40thread.v2/0?context=%7b%22Tid%22%3a%224b2a4b19-d135-420e-8bb2-b1cd238998cc%22%2c%22Oid%22%3a%22c678cf91-85c0-4c2d-82a0-cce6903f3963%22%7d

Time

2:20-3:10pm

This Friday (2/04), from 2:20-3:10pm, at the Seminar in Advances in Computing, Professor Michael Wooldridge from the University of Oxford will give a talk entitled "Artificial Intelligence – Fact and Fiction". This talk is related to his recent book "A brief history of artificial intelligence: what it is, where we are, and where we are going.""

 Abstract

Long regarded as an impossible dream, Artificial Intelligence (AI) is now an everyday reality. Advances in AI regularly make headline news, and the steady advance of AI looks set to change our world dramatically. In this lecture Professor Michael Wooldridge explores the reality of AI today: what makes AI work after half a century of effort, what is possible, and what the implications are for all of us. 

Bio

Michael Wooldridge is a Professor of Computer Science and Head of Department of Computer Science at the University of Oxford. He has been an AI researcher for more than 25 years, and has published more than 350 scientific articles on the subject. He is a Fellow of the Association for Computing Machinery (ACM), the Association for the Advancement of AI (AAAI), and the European Association for AI (EurAI). From 2014-16, he was President of the European Association for AI, and from 2015-17 he was President of the International Joint Conference on AI (IJCAI).

Location:

In person

Swearingen Engineering Center in Room 2A31

Virtual MS Teams link

Time

2:20-3:10pm

Women in Computing will be hosting its first meeting of the Spring semester at 6pm today, in Room 2277 at the Story Innovation Center! Women in Computing is open to all majors and students interesting in topics of computing technology, and diversity/inclusion within the tech industry. Everyone – all genders and majors is welcome! This meeting, we will be discussing the club, plans for this semester, and electing new members to the executive board.

If you are interest in joining, being on the executive board, and/or voting for our new members on the board please come by and see you soon! If you are unable to attend, but interested in the club, you can join our GroupMe: https://groupme.com/join_group/34681325/pIJInQ

Abstract

World models capture the dynamics of an environment and can be used to produce new, “imagined,” experiences. This can significantly reduce the number of real-world experiences required for an agent learn how to make decisions and can be combined with search for to allow an agent to plan before acting. However, in environments with a sub-symbolic representation, generating new experiences with a learned model over multiple timesteps can be difficult as small errors can accumulate over time. Furthermore, identifying previously encountered states during search can be difficult as the same state obtained by traversing different paths can result in slightly different representations. In this talk, I will discuss preliminary research on using discrete world models to address both issues. Using discrete world models, small errors can be corrected by simply rounding and identifying previously seen states is as simple as checking for equality between two arrays. Preliminary experiments with raw pixel representations of the Rubik’s cube and Sokoban show that a discrete world model can be learned using an offline dataset and can be unrolled over multiple timesteps without accumulating errors. Furthermore, after using the world model to learn a value function, combining the world model and value function with A* search solves 100% of test cases.

Bio

Forest Agostinelli is an assistant professor at the University of South Carolina. He received his B.S. from the Ohio State University, his M.S. from the University of Michigan, and his Ph.D. from the University of California, Irvine. His research group investigates how deep learning and reinforcement learning can be used to create agents that can solve complex problems and explain their solutions in a manner that humans can understand. His homepage is located at https://cse.sc.edu/~foresta/.

Location:

In person

Swearingen Engineering Center in Room 2A31

Virtual MS Teams

Friday, at the Seminar in Advances in Computing, Professor Peter Sadowski from the University of Hawai’i at Manoa will be giving a talk entitled “Deep Learning Applications in the Sciences”.

Abstract

Deep learning with artificial neural networks has enabled remarkable progress in traditional artificial intelligence applications including vision, natural language processing, and voice recognition. It also has myriad applications to science. I will review the paradigms for applying deep learning to scientific applications including inverse problems, surrogate models, and physics-informed machine learning.

Bio

Peter Sadowski is an assistant professor of Information and Computer Sciences at UH Manoa, where his lab works on a range of deep learning applications to astronomy, oceanography, and microbiome science. 

Location:

In person: Swearingen Engineering Center in Room 2A31

Virtual MS Teams

Develop collaborative assistants (chatbots) that offer innovative and ethical solutions to real-world problems !

Prizes

First Prize - $250 Second Prize - $150 Third Prize - $100

Problem ideas:

• Health: Which one or more medical specialties can treat my abdominal pain?

• Public safety: Which community is most unsafe for children?

• Water: Will there be a water problem if everyone starts washing their cars at home?

• Gardening: What happens to my water if we plant eucalyptus or cactus in SC?

• Information gathering: How is UofSC better than Clemson? Columbia v/s Charleston? based on crime statistics, hospitals, etc.

These are just suggestions to inspire. You are free to choose any problem that would help solve any problem in your community and preferably, South Carolina too.

See the event page for more information.

Meeting Location:

Storey Innovation Center 1400

Live Meeting Link for the virtual audience

Talk Abstract:  About 15 years ago, as I was working on a graphical interface for scientific software in Matlab, I got frustrated by the clumsy code structure that Matlab required for GUI coding. Although I thought that the C++ Qt library would be a great alternative, I did not want to get my fast-prototyping process slowed by low-level coding. Since Python had bindings for Qt, I decided to translate all my code into Python. To my surprise, I was able to swiftly complete this process over the weekend. Since then, I have been working almost exclusively in Python, and I never regretted it a single day. In this talk, I will the main components of the Python stack for scientific programming, focusing on neuroscience and illustrating it by summarily analyzing EEG recordings (MNE-Python). I will discuss why Python has become a major player in this field and how limitations typical to interpreted languages (e.g., slow at runtime) have been tackled with libraries such as NumPy. I will also explain why Python is a strong environment for data wrangling by introducing libraries like Pandas – which offers data frame functionalities similar to R – and XArray. Finally, I will touch upon how libraries like Seaborn provide a high-level interface for quickly producing publication-quality figures with only a few (if not a single) lines of code.

Speaker's Bio: Christian O’Reilly received his B.Ing (elec eng; 2007), his M.Sc.A. (biomed eng; 2011), and his Ph.D. (biomed eng; 2012) from Polytechnique Montreal. He was a postdoc fellow at the CARSM (2012-2014) and then a NSERC postdoc fellow at McGill's Brain Imaging Center (2014-2015) where he worked on EEG sleep transients. He also worked at the EPFL (2015-2018) on modeling of the thalamocortical loop and at McGIll on brain connectivity (2020-2021). Since 2021, he is Assistant Professor at UofSC.

Meeting Location:

Storey Innovation Center 1400

Live Meeting Link for the virtual audience :

https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWZkMTYwZDQtMDVmMy00NjA1LTgxNmEtZDExMDdiZTM2ZjYz%40thread.v2/0?context=%7b%22Tid%22%3a%224b2a4b19-d135-420e-8bb2-b1cd238998cc%22%2c%22Oid%22%3a%225fc2170a-7068-4a33-9021-df11b94ba696%22%7d

Talk Abstract: David’s talk will cover some of the exciting technologies the Devices, Circuits & Systems group at ARM is researching as well as what he sees in the general trend and future of process technology. And since it’s not possible to discuss CMOS scaling without commenting on Moore’s law, he will do that, too. 🙂

Speaker's Bio: David Pietromonaco has been in the semiconductor industry for almost 30 years at Hewlett-Packard, Sony, and most recently Artisan/Arm (for 20 of those). He works in Arm Research; in the Devices, Circuits & Systems group, specifically on the Technology Optimized Design team. That team tries to look 5-10 years ahead to understand future computing technologies and how to utilize them.

DISSERTATION DEFENSE

                                                                                  Department of Computer Science and Engineering

University of South Carolina 

Author : John Ravan

Advisor : Dr. Csilla Farkas

Date : November 1, 2021

Time : 1:30pm

Place : Virtual Defense

Join Zoom Meeting

https://citadelonline.zoom.us/j/5257755660?pwd=dzBwNW85RUdSRjVWdGp4RzRxbzE2UT09

Abstract

Concurrent database transactions within a web service environment can cause a variety of problems without the proper concurrency control mechanisms in place. A few of these problems involve data integrity issues, deadlock, and efficiency issues. Even with today's industry standard solutions to these problems, they have taken a reactive approach rather than proactively preventing these problems from happening. We deliver a twofold solution that presents a proactive prediction-based approach to ensure consistency while keeping execution time the same or faster than current industry solutions. The first part of this solution involves prototyping and formally proving a prediction-based scheduler. 

The prediction-based scheduler leverages a prediction-based metric that promotes transactions with reliable reputations based on the transaction's performance metric. This performance metric is based on the transaction's likelihood to commit and its efficiency within the system. We can then predict the outcome of the transaction based on the metric and apply customized lock behaviors to address consistency issues in current web service environments. We have formally proven that the solution will increase consistency among web service transactions without a performance degradation that is worse than industry standard 2PL. The simulation was developed using a multi-threaded approach to simulate concurrent transactions.  Experimentation results show that the solution works comparatively with industry solutions with the added benefit of ensured consistency in some cases and deadlock avoidance in others. This work has been published in IEEE Transactions on Services Computing. 

The second part of the solution involves building the prediction-based metric mentioned previously. In the initial solution we assumed the prediction-based categorization coming into the solution in order to prove the feasibility and correctness of a prediction-based scheduler. 

Once that was established, we extended the four-category solution to a dynamic reputation score built upon transactional attributes. The attributes used in the reputation score are system abort ranking, user abort ranking, efficiency ranking, and commit ranking. With these four attributes we were able to establish a dynamic dominance structure that allowed for a transaction to promote or demote itself based on its performance within the system. This work has been submitted to ACM Transactions on Information Systems and awaiting review. 

Both phases provide a complete solution of prediction-based transaction scheduling that provides dynamic categorization no matter the transactional environment. 

Future work of this system would involve extending the prediction-based solution to a multi-level secure database with an added dimension. The dimension provides a security classification in addition to attributes for dynamic reputation that allows for transactions to establish dominance. The goal would be to prevent covert timing channels that occur in multi-level secure database systems due to the differing classifications. Our reputation score would provide a cover story for timing differences of transactions of different security levels to allow for a more robust scheduling algorithm. This would allow for high security transactions to gain priority over low security transactions without exposing a covert timing channel.