Defense Notices

Scene understanding in both static images and dynamic videos is the ultimate goal in computer vision. As two important sub-tasks of this endeavor, object detection and tracking have been extensively studied in the past decades, however, the problem is still not well addressed. The main challenge is that the appearance of objects is affected by a number of factors, such as scale, occlusion, illumination, and so on. Recently, deep learning has attracted lots of interests in the computer vision community. However, how to tackle these challenges in object detection and tracking is still an open problem. In this work, we propose a method for detecting objects in images using a single deep neural network, which can be optimized end-to-end and predict the object bounding boxes and class probabilities in one evaluation. To handle the challenges in object tracking, we propose a framework, which consists of a novel deep Convolutional Neural Networks (CNNs) to effectively generate robust spatial appearance, and a Long Short-term Memory (LSTM) network that incorporates temporal information to achieve long-term object tracking accuracy in real-time.

According to Moore’s law the number of transistors per square inch double every two years. Scaling down technology reduces size and cost however, also increases the number of problems. Our current computers using Von-Neumann architectures are seeing progressive difficulties not only due to scaling down the technology but also due to grid-lock situation in its architecture. As a solution to this, scientists came up architectures whose function resembles that of the brain. They called these brains inspired architectures, neuromorphic computers. The building block of the brain is the neuron which encodes, decodes and processes the data. The neuron is known to accept sensory information and converts this information into a spike train. This spike train is encoded by the neuron using different ways depending on the situation. Rate encoding, temporal encoding, population encoding, sparse encoding and rate-order encoding are a few encoding schemes said to be used by the neuron. These different neural encoding schemes are discussed as the primary focus of the thesis. A comparison between these different schemes is also provided for better understanding, thus helping in the design of an efficient neuromorphic computer. This thesis also focusses on hardware implementation of a neuron. Leaky Fire and Integrate neuron model has been used in this work which uses spike-time dependent encoding. Different neuron models are discussed with a comparison as to which model is effective under which circumstances. The electronic neuron model was implemented using 180nm CMOS Technology using Global Foundries PDK libraries. Simulation results for the neuron are presented for different inputs and different excitation currents. These results show the successful encoding of sensory information into a spike train.

Due to the increasing need for greater Radio Frequency (RF) spectrum by mobile apps like Facebook and Instagram, high data-rate communication protocols like 5G and the Internet of Things, it has led to the issue of spectrum congestion as radar systems have traditionally maintain the largest share of the RF spectrum. To resolve the spectrum congestion problem, it has become even necessary for users from both types of systems to coexist within a finite spectrum allocation. However, this then leads to other problems such as the increased likelihood of mutual interference experienced by all users that are coexisting within the finite spectrum. 

In this dissertation, we propose to address the problem of spectrum congestion via a two-step approach. The first step of this approach involves designing an optimal sparse spectrum allocation scheme to radar systems such that the radar range resolution performance can be maintained with a smaller resulting bandwidth at a cost of degraded sidelobe performance. The second step of this approach involves designing radar waveforms that possesses good spectral containment property by expanding the framework of Polyphase-coded Frequency Modulated (PCFM) waveforms to higher-order representations such that these waveforms will mitigate issues of interference experienced by other systems when both systems are coexisting within the same band. 

Von Neumann Bottleneck, which refers to the limited throughput between the CPU and memory, has already become the major factor hindering the technical advances of computing systems. In recent years, neuromorphic systems started to gain the increasing attentions as compact and energy-efficient computing platforms. As one of the most crucial components in the neuromorphic computing systems, neural encoder transforms the stimulus (input signals) into spike trains. In this report, I will present my research work on spike-time-dependent encoding schemes and its relevant energy efficient encoders’ design. The performance comparison among rate encoding, latency encoding, and temporal encoding would be discussed in this report. The proposed neural temporal encoder allows efficient mapping of signal amplitude information into a spike time sequence that represents the input data and offers perfect recovery for band-limited stimuli. The simulation and measurement results show that the proposed temporal encoder is proven to be robust and error-tolerant. 

With an increasing trend of adoption of machine learning in various real-world problems, the need for transparent and reliable models has become apparent. Especially in some socially consequential applications, such as medical diagnosis, credit scoring, and decision making in educational systems, it may be problematic if humans cannot understand and trust those models. To this end, in this work, we propose a novel machine learning algorithm, constructivism learning. To achieve transparency, we formalized a Bayesian nonparametric approach using sequential Dirichlet Process Mixture of prediction models to support constructivism learning. To achieve reliability, we exploit two strategies, reducing model uncertainty and increasing task construction stability by leveraging techniques in active learning and self-paced learning. 

Distance metrics are concerned with learning how objects are similar, and are a critical component of many machine learning algorithms such as k-nearest neighbors and kernel machines. Traditional metrics are unable to adapt to data with heterogenous interactions in the feature space. State of the art methods consider learning multiple metrics, each in some way local to a portion of the data. Selecting how the distance metrics are local to the data is done apriori, with no known best approach. 
In this proposal, we address the local metric learning scenario from three complementary perspectives. In the first direction, we consider a spatial approach, and develop an efficient Frank-Wolfe based technique to learn local distance metrics directly in a high-dimensional input space. We then consider a view-local perspective, where we associate each metric with a separate view of the data, and show how the approach naturally evolves into a multiple kernel learning problem. Finally, we propose a new function for learning a metric which is based on a newly discovered operator called the t-product, here we show that our metric is composed of multiple parts, with each portion local to different interactions in the input space. 

Resource limited embedded systems provide a great challenge to programming using functional languages. Although we cannot program these embedded systems directly with Haskell, we show than an embedded domain specific language is able to be used to program them, providing a user friendly environment for both prototyping and full development. The Arduino line of microcontroller boards provide a versatile, low cost and popular platform for development of these resource limited systems, and we use this as the platform for our DSL research. 

First we provide a shallowly embedded domain specific language and a firmware interpreter, allowing the user to program the Arduino while tethered to a host computer. Second, we add a deeply embedded version, allowing the interpreter to run standalone from the host computer, as well as allowing us to compile the code to C and then machine code for efficient operation. Finally, we develop a method of transforming the shallowly embedded DSL syntax into the deeply embedded DSL syntax automatically.

With the promise of meeting future capacity demands for mobile broadband communications, 3D massive-MIMO/Full Dimension MIMO (FD-MIMO) systems have gained much interest among the researchers in recent years. Apart from the huge spectral efficiency gain offered by the system, the reason for this great interest can also be attributed to significant reduction of latency, simplified multiple access layer, and robustness to interference. However, in order to completely extract the benefits of massive-MIMO systems, accurate channel state information is critical. In this thesis, a channel estimation method based on direction of arrival (DoA) estimation is presented for massive- MIMO OFDM systems. To be specific, the DoA is estimated using Estimation of Signal Parameter via Rotational Invariance Technique (ESPRIT) method, and the root mean square error (RMSE) of the DoA estimation is analytically characterized for the corresponding MIMO-OFDM system.

Billions of dollars are lost every year to successful cyber attacks that are fundamentally enabled by software vulnerabilities. Modern cyber attacks increasingly threaten individuals, organizations, and governments, causing service disruption, inconvenience, and costly incident response. Given that such attacks are primarily enabled by software vulnerabilities, this work examines the efficacy of using change metrics, along with architectural burst and maintainability metrics, to predict modules and files that should be analyzed or tested further to excise vulnerabilities prior to release. 

The problem addressed by this research is the residual vulnerability problem, or vulnerabilities that evade detection and persist in released software. Many modern software projects are over a million lines of code, and composed of reused components of varying maturity. The sheer size of modern software, along with the reuse of existing open source modules, complicates the questions of where to look, and in what order to look, for residual vulnerabilities. 

Traditional code complexity metrics, along with newer frequency based churn metrics (mined from software repository change history), are selected specifically for their relevance to the residual vulnerability problem. We compare the performance of these complexity and churn metrics to architectural level change burst metrics, automatically mined from the git repositories of the Mozilla Firefox Web Browser, Apache HTTP Web Server, and the MySQL Database Server, for the purpose of predicting attack prone files and modules. 

We offer new empirical data quantifying the relationship between our selected metrics and the severity of vulnerable files and modules, assessed using severity data compiled from the NIST National Vulnerability Database, and cross-referenced to our study subjects using unique identifers defined by the Common Vulnerabilities and Exposures (CVE) vulnerability catalog. Our results show that architectural level change burst metrics can perform well in situations where more traditional complexity metrics fail as reliable estimators of vulnerability severity. In particular, results from our experiments on Apache HTTP Web Server indicate that architectural level change burst metrics show high correlation with the severity of known vulnerable modules, and do so with information directly available from the version control repository change-set (i.e., commit) history. 

Incomplete and numerical data are common for many application domains. There have been many approaches to handle missing data in statistical analysis and data mining. To deal with numerical data, discretization is crucial for many machine learning algorithms. However, most of the discretization algorithms cannot be applied to incomplete data sets. 

Multiple Scanning is an entropy based discretization method. Previous research shown it outperforms commonly used discretization methods: Equal Width or Equal Frequency discretization. In this work, Multiple Scanning is tested on C4.5 and MLEM2 on incomplete datasets. Results show for some data sets, the setup utilizing Multiple Scanning as preprocessing performs better, for the other data sets, C4.5 or MLEM2 should be used by themselves. Our conclusion is that there are no universal optimal solutions for all data sets. Setup should be custom-made.