Graduate defenses this upcoming week

Graduate Defenses
Graduate Defenses

Ph.D. Thesis Defense: Guolong Zheng
Thursday, April 21, 2022
11:00 AM (CST)
Join Zoom Meeting

“Ensure Correctness For Imperative and Declarative Programs”

There are two different types of programs: imperative program that describes how to solve a problem and declarative program that describes how to recognize that the problem is solved. The inherent difference between them leads to different analysis techniques. In this dissertation, we develop tools to ensure correctness and improve reliability for both imperative and declarative programs. For imperative program, we focus on automatically infer program invariant. Program invariant, which describes program states or behaviors as mathematical formulas, is one main approach to verify and debug imperative programs. However, current invariant generation techniques focus on numerical invariant described by Hoare Logic, lacking support for memory related properties. We present SLING, a dynamic analysis tool to automatically infer program invariant described by Separation Logic, an extension of Hoare logic with compact description for memory properties. The empirical results show that SLING can efficiently discover invariant for pointers and data structures and the generated invariant can be used automatically repair corrupted data structures. For declarative program, we focus on automatically debugging Alloy specifications. Alloy is a declarative specification language widely used in various software problems. However, unlike imperative programs, there is a dearth of techniques to help debugging Alloy specifications. We present FLACK to automatically locate Alloy bugs and ATR to repair them. Experimental results show that FLACK can accurately locate buggy expressions and ATR can effectively repair different types of bugs.

Committe Members:
ThanhVu Nguyen, Advisor
Hamid Bagheri, Co-Advisor
Lisong Xu
Matthew Dwyer
Shane Farritor

Ph.D. Dissertation Defense: Rubi Quiñones
Friday, April 22, 2022
9:00 AM (CST)
111 Avery Hall
Join Zoom Meeting:

"Unsupervised Cosegmentation and Phenotypes for Multi -Modal, -Perspective, and -State Imagery"

Cosegmentation is used to segment the object(s) from the background by simultaneously analyzing multiple images. The state-of-the-art cosegmentation methods do not satisfactorily address the challenges in segmenting multiple images of evolving objects such as growth sequences of plants. Segmenting the plant from the background is critical in plant phenotyping applications whose goal is to compute any measurable traits from plant images. Current segmentation algorithms typically use low-level methods to isolate the plant from the background. The results are often inadequate leading to degradation in the performance of subsequent algorithms, including the computation of phenotypes. The goal of this research is to increase segmentation accuracy in the plant phenotyping domain by leveraging cosegmentation techniques. First, we show the significant bias in current cosegmentation algorithms and datasets by testing their efficacy of a multiple-aspects (views, modalities, and time) plant imagery dataset. Second, we develop an unsupervised cosegmentation-coattention deep learning framework, called OSC-CO2, that can leverage the results of multiple algorithms specifically meant to increase the overall accuracy in a multiple-aspects plant image dataset. Finally, we propose a suite of novel phenotypes that leverage the information from multiple images across an aspect to describe information-rich phenotypes that represent higher-order properties of the plant. The efficacy of the segmentation algorithm and the phenotypes is demonstrated using multi-aspect imagery from a high-throughput plant phenotyping facility. The broader impact of this research is the advancement of plant phenomics in understanding the plant’s environmental interactions for maximal resilience and yield and indirectly aiding food security and sustainable crop production.

Committee Members:
Ashok Samal, Chair
Sruti Das Choudhury, Co-Chair
Francisco Munoz-Arriola, Co-Chair
HongFeng Yu, Member & Reader
Christopher Bohn, Member
Tala Awada, Outside Representative & Reader

Ph.D. Dissertation Defense: Yuji Mo
Monday, April 25, 2022
3:00 PM (CST)
Via Zoom:

“Cost-Aware Hierarchical Active Learning and Sub-Linear Time Embedding Based Deep Retrieval”

The research in this dissertation consists of two parts: An active learning algorithm for hierarchical labels and an embedding based retrieval algorithm. In the first part, we present a new active learning model that works with training data labeled according to a hierarchical scheme, e.g., data from named entity recognition (NER), document classification, and biological sequence analysis. Our model allows a high-level learning problem (e.g., location versus non-location) to be decomposed into a collection of finer-grained problems (e.g., museum versus non-museum), which can then be learned individually and the results combined as an alternative means to learn the high-level concept. We show that this approach can result in higher levels of precision for the same level of recall. Since such finer-grained labeling data could be more expensive to obtain, we work in the active learning setting and study the trade-off between the increased cost of purchasing finer-grained labels versus the potential increased benefit in learning. We then present a family of parameterized algorithms to work in our new model, and empirically evaluate our approach on NER and document classification problems.

In the second part, we present a Bayesian Deep Structured Semantic Model (BDSSM) that efficiently in retrieval tasks with a large pool of candidates for real time applications, e.g., in search engines, digital ads and recommendation systems. The efficiency is achieved by indexing the items into groups based on their sparse representation of embeddings during offline pre-computation. In the online retrieval phase, the algorithm only retrieves and ranks items from indices that are relevant to the query. We explore optimization strategies in the algorithm to make sparse representation sparser. In evaluation, the algorithm is compared it with other popular clustering-based, hashing-based and tree-based retrieval methods. We measure the differences in multiple dimensions, including retrieval recall, storage of embeddings, and CPU time. We show that this algorithm outperforms other algorithms in the comparison in both recall and CPU time with the same storage limit. Finally, we also show that this algorithm can be used in exploration when the model is recurrently retrained.

Dr. Stephen Scott, Advisor
Dr. Ashok Samal
Dr. Leen-Kiat Soh
Dr. Etsuko Moriyama

M.S. Thesis Defense: Kaustubh Gupta
Tuesday, April 26, 2022
11:00 AM (CST)
Via Zoom
Join Zoom Meeting:

"Machine Learning Based Device Type Classification for IoT device Continuous and Re-Authentication"

Today, the use of Internet of Things (IoT) devices is higher than ever and it is growing rapidly. Many IoT devices are usually manufactured by home appliance manufacturers where the security and the privacy is not the foremost concern. When an IoT device is connected to a network, currently, there does not exist a strict authentication method that verifies the identity of the device, allowing any rogue IoT device to authenticate to an access point. This thesis addresses the issue by introducing methods for continuous and re-authentication of static and dynamic IoT devices respectively. We introduce mechanisms and protocols for authenticating a device in a network through leveraging Machine Learning (ML) to classify not only if the device is IoT or not but also the type of IoT device attempting to connect to the network with an accuracy over 95%. Furthermore, we compare different types of machine learning classifiers to best estimate the types of IoT device and use them to develop a stricter and more efficient method for authentication.

Dr. Nirnimesh Ghose,
Dr. Byrav Ramamurthy,
Dr. Lisong Xu

Ph.D. Dissertation Defense: Tian Gao
Tuesday, April 26, 2022
2:30 PM (CST)
Via Zoom:

"Imaging Systems And Data Analysis for Plant Phenotyping"

High-throughput phenotyping is becoming a critical method for many plant science researchers. Compared to manual phenotyping, high-throughput phenotyping shows advantages in the efficiency and the capability of large-scale measurements. One of the most popular existing solutions for high-throughput phenotyping is RGB image-based methods for easy access. However, traditional RGB image-based methods suffer from inaccuracy in two aspects. First, the image-based methods will inevitably lose the depth information as images are projections of 3D objects onto 2D planes. Second, the RGB image-based methods only involve three bands (i.e., Red, Green, and Blue), and information contained in other bands is lost in image capturing. To solve the bottleneck of RGB image-based methods, I have developed end-to-end solutions to tackle large-scale 3D-based data and hyperspectral data. First, I design multiple imaging systems for plant phenotyping and an end-to-end pipeline to generate the denoised point clouds. By comparing the ground truth with the extracted results, I provide insights into the performance and present evidence for increased accuracy from the imaging systems. Moreover, I present a systematic analysis of how different settings of the imaging systems affect the results. Second, using the point clouds generated by our imaging systems and pipeline, I design a method to detect the critical events on both whole plants and specific components. I implement an experiment on maize for evaluation and successfully detect events in the process of plant growth. Third, I develop a novel end-to-end platform to provide hyperspectral information for seeds, including a high-throughput imaging system and open-source software. To fully demonstrate the platform's effectiveness, I conduct a classification of seeds using machine learning models. Moreover, my experiment has shown the potential for seed segmentation with different seed species.

Committee members:
Hongfeng Yu (Chair of Committee)
Juan Cui (Reader)
Lisong Xu (Reader)
Harkamal Walia (Outside Representative)