Recent PhD theses

Perreault, Logan Jared. (Montana State University) “On the Usability of Continuous Time Bayesian Networks: Improving Scalability and Expressiveness” (2017), Advisor: Sheppard, John W., https://www.cs.montana.edu/sheppard/

The Continuous Time Bayesian Network (CTBN) is a model capable of compactly representing the behavior of discrete state systems that evolve in continuous time. This is achieved by factoring a Continuous Time Markov Process using the structure of a directed graph. Although CTBNs have proven themselves useful in a variety of applications, adoption of the model for use in real-world problems can be difficult. The authors believe this is due in part to limitations relating to scalability as well as representational power and ease of use. This dissertation attempts to address these issues.

First, the authors improve the expressiveness of CTBNs by providing procedures that support the representation of non-exponential parametric distributions. The authors also propose the Continuous Time Decision Network (CTDN) as a framework for representing decision problems using CTBNs. This new model supports optimization of a utility value as a function of a set of possible decisions. Next, the authors address the issue of scalability by providing two distinct methods for compactly representing CTBNs by taking advantage of similarities in the model parameters. These compact representations are able to mitigate the exponential growth in parameters that CTBNs exhibit, allowing for the representation of more complex processes. The authors then introduce another approach to managing CTBN model complexity by introducing the concept of disjunctive interaction for CTBNs. Disjunctive interaction has been used in Bayesian networks to provide significant reductions in the number of parameters, and the authors have adapted this concept to provide the same benefits within the CTBN framework.

Finally, the authors demonstrate how CTBNs can be applied to the real-world task of system prognostics and diagnostics. The authors show how models can be built and parameterized directly using information that is readily available for diagnostic models. The authors then apply these model construction techniques to build a CTBN describing a vehicle system. The vehicle model makes use of some of the newly introduced algorithms and techniques, including the CTDN framework and disjunctive interaction. This extended application not only demonstrates the utility of the novel contributions presented in this work, but also serves as a template for applying CTBNs to other real-world problems.

 
Aditham, Santosh. University of South Florida, “Mitigation of Insider Attacks for Data Security in Distributed Computing Environments” (2017), Advisor: Ranganathan, Nagarajan http://www.usf.edu/engineering/cse/people/ranganathan-nagarajan.aspx and Katkoori, Srinivas http://www.usf.edu/engineering/cse/people/katkoori-srinivas.aspx

In big data systems, the infrastructure is such that large amounts of data are hosted away from the users. Information security is a major challenge in such systems. From the customer’s perspective, one of the big risks in adopting big data systems is in trusting the service provider who designs and owns the infrastructure, with data security and privacy. However, big data frameworks typically focus on performance and the opportunity for including enhanced security measures is limited. In this dissertation, the problem of mitigating insider attacks is extensively investigated and several static and dynamic run-time techniques are developed. The proposed techniques are targeted at big data systems but applicable to any data system in general. First, a framework is developed to host the proposed security techniques and integrate with the underlying distributed computing environment. The authors endorse the idea of deploying this framework on special purpose hardware and a basic model of the software architecture for such security coprocessors is presented. Then, a set of compile-time and run-time techniques are proposed to protect user data from the perpetrators. These techniques target detection of insider attacks that exploit data and infrastructure. The compile-time intrusion detection techniques analyze the control flow by disassembling program binaries while the run-time techniques analyze the memory access patterns of processes running on the system. The proposed techniques have been implemented as prototypes and extensively tested using big data applications. Experiments were conducted on big data frameworks such as Hadoop and Spark using cloud-based services. Experimental results indicate that the proposed techniques successfully detect insider attacks in the context of data loss, data degradation, data exposure and infrastructure degradation.

 
Steinemann, Natalie Anna. The City College of New York, “Perceptual decision making in humans: Neural correlates along the sensorimotor hierarchy” (2017) Advisor: Kelly, Simon P. https://www.gc.cuny.edu/Page-Elements/Academics-Research-Centers-Initiat...

Decision-making has been studied in a variety of fields ranging from economics to computer science. Within the neural sciences, a perceptual decision has been defined as the translation of sensory perception into goal-directed actions. Great strides have been achieved in uncovering the neural underpinnings of decision-making primarily through single-unit recordings in non-human primates. The results of these studies suggest that decisions are computed in neural populations specific to the effector employed to report the outcome of the decision. Results obtained by non-invasive measures in humans, on the other hand, indicate that a supra-modal decision area may exist, which drives the activity observed in effector-specific neural populations. Due to differences in the scale of measurement, it has, however, been challenging to translate these groundbreaking findings to the human brain where data recordings are commonly limited to non-invasive electrophysiology and neuroimaging.
The research summarized in this dissertation was designed to gain a more complete understanding of the functional significance of decision correlates measured non-invasively in humans. The authors first characterize non-invasively measured signals of sensory representation, decision formation, motor preparation, and response execution, and give new insight into the differential influence of evidence and response time constraints all of these key stages. In line with previous reports in human and non-human primates, the authors found response time constraints to have a substantial effect on signatures of motor preparation before the onset of sensory evidence as well as during evidence evaluation. Additionally, the authors found that response time constraints enhance the differential representation of sensory evidence and accelerate the process of motor response execution. In a second study, the authors use neuroimaging to localize potential sources of effector-independent evidence accumulation in the human brain. Neural populations in the supramarginal gyrus of the parietal lobe, right operculum, left insula, and left inferior frontal gyrus are identified as candidate regions. Among the four regions, it is suggested that the neural population in supramarginal gyrus may be situated upstream of the other three in the sensorimotor hierarchy. Finally, the authors identify the means by which non-invasive decision correlates measured in humans may advance the understanding of other cognitive processes such as memory formation and aid in the diagnosis of learning disorders.
 
Wang, Panqu. University of California, San Diego, “Towards The Deep Model: Understanding Visual Recognition Through Computational Models”, (2017), Advisor: Cottrell, Garrison https://cseweb.ucsd.edu/~gary/ and Vasconcelos, Nuno http://www.svcl.ucsd.edu/~nuno/

Understanding how visual recognition is achieved in the human brain is one of the most fundamental questions in vision research. In this thesis the authors seek to tackle this problem from a neurocomputational modeling perspective. More specifically, the authors build machine learning-based models to simulate and explain cognitive phenomena related to human visual recognition, and the authors improve computational models using brain-inspired principles to excel at computer vision tasks. The authors first describe how a neurocomputational model (“The Model”, TM, (Cottrell & Hsiao, 2011)) can be applied to explain the modulation of visual experience on the performance of subordinate-level face and object recognition. Next, by introducing a mixture-of-experts structure in the model, the authors show that TM can be used to simulate the development of hemispheric lateralization of face processing. In addition, the authors extend TM to “The Deep Model” (TDM) by coupling it with deep learning techniques, and use TDM to explain the peripheral vision advantage in human scene recognition. Furthermore, the authors show the performance of these computational methods can be improved by introducing realistic constraints based on the human brain. By combining unsupervised feature learning principles with the Gnostic Fields theory of how the brain performs object recognition across the ventral visual pathway, the authors show a biologically-inspired model can develop realistic features of the early visual cortex, while performing well on object recognition datasets. By designing better encoding and decoding strategies in the deep neural network, the authors demonstrate that their system achieves the state-of-the-art performance on pixel-level semantic segmentation task on many popular computer vision benchmarks.