Imagery data has become important for civil infrastructure operation and

maintenance because imagery data can capture detailed visual information with high

frequencies. Computer vision can be useful for acquiring spatiotemporal details to

support the timely maintenance of critical civil infrastructures that serve society. Some

examples include: irrigation canals need to maintain the leaking sections to avoid water

loss; project engineers need to identify the deviating parts of the workflow to have the

project finished on time and within budget; detecting abnormal behaviors of air traffic

controllers is necessary to reduce operational errors and avoid air traffic accidents.

Identifying the outliers of the civil infrastructure can help engineers focus on targeted

areas. However, large amounts of imagery data bring the difficulty of information

overloading. Anomaly detection combined with contextual knowledge could help address

such information overloading to support the operation and maintenance of civil

infrastructures.

Some challenges make such identification of anomalies difficult. The first challenge is

that diverse large civil infrastructures span among various geospatial environments so

that previous algorithms cannot handle anomaly detection of civil infrastructures in

different environments. The second challenge is that the crowded and rapidly changing

workspaces can cause difficulties for the reliable detection of deviating parts of the

workflow. The third challenge is that limited studies examined how to detect abnormal

behaviors for diverse people in a real-time and non-intrusive manner. Using video andii

relevant data sources (e.g., biometric and communication data) could be promising but

still need a baseline of normal behaviors for outlier detection.

This dissertation presents an anomaly detection framework that uses contextual

knowledge, contextual information, and contextual data for filtering visual information

extracted by computer vision techniques (ADCV) to address the challenges described

above. The framework categorizes the anomaly detection of civil infrastructures into two

categories: with and without a baseline of normal events. The author uses three case

studies to illustrate how the developed approaches can address ADCV challenges in

different categories of anomaly detection. Detailed data collection and experiments

validate the developed ADCV approaches.

Robot motion planning requires computing a sequence of waypoints from an initial configuration of the robot to the goal configuration. Solving a motion planning problem optimally is proven to be NP-Complete. Sampling-based motion planners efficiently compute an approximation of the optimal solution. They sample the configuration space uniformly and hence fail to sample regions of the environment that have narrow passages or pinch points. These critical regions are analogous to landmarks from planning literature as the robot is required to pass through them to reach the goal.

This work proposes a deep learning approach that identifies critical regions in the environment and learns a sampling distribution to effectively sample them in high dimensional configuration spaces.

A classification-based approach is used to learn the distributions. The robot degrees of freedom (DOF) limits are binned and a distribution is generated from sampling motion plan solutions. Conditional information like goal configuration and robot location encoded in the network inputs showcase the network learning to bias the identified critical regions towards the goal configuration. Empirical evaluations are performed against the state of the art sampling-based motion planners on a variety of tasks requiring the robot to pass through critical regions. An empirical analysis of robotic systems with three to eight degrees of freedom indicates that this approach effectively improves planning performance.

Referring Expression Comprehension (REC) is an important area of research in Natural Language Processing (NLP) and vision domain. It involves locating an object in an image described by a natural language referring expression. This task requires information from both Natural Language and Vision aspect. The task is compositional in nature as it requires visual reasoning as underlying process along with relationships among the objects in the image. Recent works based on modular networks have

displayed to be an effective framework for performing visual reasoning task.

Although this approach is effective, it has been established that the current benchmark datasets for referring expression comprehension suffer from bias. Recent work on CLEVR-Ref+ dataset deals with bias issues by constructing a synthetic dataset

and provides an approach for the aforementioned task which performed better than the previous state-of-the-art models as well as showing the reasoning process. This work aims to improve the performance on CLEVR-Ref+ dataset and achieve comparable interpretability. In this work, the neural module network approach with the attention map technique is employed. The neural module network is composed of the primitive operation modules which are specific to their functions and the output is generated using a separate segmentation module. From empirical results, it is clear that this approach is performing significantly better than the current State-of-theart in one aspect (Predicted programs) and achieving comparable results for another aspect (Ground truth programs)

Hyperbolic geometry, which is a geometry which concerns itself with hyperbolic space, has caught the eye of certain circles in the machine learning community as of late. Lauded for its ability to encapsulate strong clustering as well as latent hierarchies in complex and social networks, hyperbolic geometry has proven itself to be an enduring presence in the network science community throughout the 2010s, with no signs of fading into obscurity anytime soon. Hyperbolic embeddings, which map a given graph to hyperbolic space, have particularly proven to be a powerful and dynamic tool for studying complex networks. Hyperbolic embeddings are exploited in this thesis to illustrate centrality in a graph. In network science, centrality quantifies the influence of individual nodes in a graph. Eigenvector centrality is one type of such measure, and assigns an influence weight to each node in a graph by solving for an eigenvector equation. A procedure is defined to embed a given network in a model of hyperbolic space, known as the Poincare disk, according to the influence weights computed by three eigenvector centrality measures: the PageRank algorithm, the Hyperlink-Induced Topic Search (HITS) algorithm, and the Pinski-Narin algorithm. The resulting embeddings are shown to accurately and meaningfully reflect each node's influence and proximity to influential nodes.

Large organizations have multiple networks that are subject to attacks, which can be detected by continuous monitoring and analyzing the network traffic by Intrusion Detection Systems. Collaborative Intrusion Detection Systems (CIDS) are used for efficient detection of distributed attacks by having a global view of the traffic events in large networks. However, CIDS are vulnerable to internal attacks, and these internal attacks decrease the mutual trust among the nodes in CIDS required for sharing of critical and sensitive alert data in CIDS. Without the data sharing, the nodes of CIDS cannot collaborate efficiently to form a comprehensive view of events in the networks monitored to detect distributed attacks. The compromised nodes will further decrease the accuracy of CIDS by generating false positives and false negatives of the traffic event classifications. In this thesis, an approach based on a trust score system is presented to detect and suspend the compromised nodes in CIDS to improve the trust among the nodes for efficient collaboration. This trust score-based approach is implemented as a consensus model on a private blockchain because private blockchain has the features to address the accountability, integrity and privacy requirements of CIDS. In this approach, the trust scores of malicious nodes are decreased with every reported false negative or false positive of the traffic event classifications. When the trust scores of any node falls below a threshold, the node is identified as compromised and suspended. The approach is evaluated for the accuracy of identifying malicious nodes in CIDS.

Languages, specially gestural and sign languages, are best learned in immersive environments with rich feedback. Computer-Aided Language Learning (CALL) solu- tions for spoken languages have successfully incorporated some feedback mechanisms, but no such solution exists for signed languages. Computer Aided Sign Language Learning (CASLL) is a recent and promising field of research which is made feasible by advances in Computer Vision and Sign Language Recognition(SLR). Leveraging existing SLR systems for feedback based learning is not feasible because their decision processes are not human interpretable and do not facilitate conceptual feedback to learners. Thus, fundamental research is needed towards designing systems that are modular and explainable. The explanations from these systems can then be used to produce feedback to aid in the learning process.

In this work, I present novel approaches for the recognition of location, movement and handshape that are components of American Sign Language (ASL) using both wrist-worn sensors as well as webcams. Finally, I present Learn2Sign(L2S), a chat- bot based AI tutor that can provide fine-grained conceptual feedback to learners of ASL using the modular recognition approaches. L2S is designed to provide feedback directly relating to the fundamental concepts of ASL using an explainable AI. I present the system performance results in terms of Precision, Recall and F-1 scores as well as validation results towards the learning outcomes of users. Both retention and execution tests for 26 participants for 14 different ASL words learned using learn2sign is presented. Finally, I also present the results of a post-usage usability survey for all the participants. In this work, I found that learners who received live feedback on their executions improved their execution as well as retention performances. The average increase in execution performance was 28% points and that for retention was 4% points.

Internet memes have become a widespread tool used by people for interacting and exchanging ideas over social media, blogs, and open messengers. Internet memes most commonly take the form of an image which is a combination of image, text, and humor, making them a powerful tool to deliver information. Image memes are used in viral marketing and mass advertising to propagate any ideas ranging from simple commercials to those that can cause changes and development in the social structures like countering hate speech.

This work proposes to treat automatic image meme generation as a translation process, and further present an end to end neural and probabilistic approach to generate an image-based meme for any given sentence using an encoder-decoder architecture. For a given input sentence, a meme is generated by combining a meme template image and a text caption where the meme template image is selected from a set of popular candidates using a selection module and the meme caption is generated by an encoder-decoder model. An encoder is used to map the selected meme template and the input sentence into a meme embedding space and then a decoder is used to decode the meme caption from the meme embedding space. The generated natural language caption is conditioned on the input sentence and the selected meme template.

The model learns the dependencies between the meme captions and the meme template images and generates new memes using the learned dependencies. The quality of the generated captions and the generated memes is evaluated through both automated metrics and human evaluation. An experiment is designed to score how well the generated memes can represent popular tweets from Twitter conversations. Experiments on Twitter data show the efficacy of the model in generating memes capable of representing a sentence in online social interaction.

There have been multiple attempts of coupling neural networks with external memory components for sequence learning problems. Such architectures have demonstrated success in algorithmic, sequence transduction, question-answering and reinforcement learning tasks. Most notable of these attempts is the Neural Turing Machine (NTM), which is an implementation of the Turing Machine with a neural network controller that interacts with a continuous memory. Although the architecture is Turing complete and hence, universally computational, it has seen limited success with complex real-world tasks.

In this thesis, I introduce an extension of the Neural Turing Machine, the Neural Harvard Machine, that implements a fully differentiable Harvard Machine framework with a feed-forward neural network controller. Unlike the NTM, it has two different memories - a read-only program memory and a read-write data memory. A sufficiently complex task is divided into smaller, simpler sub-tasks and the program memory stores parameters of pre-trained networks trained on these sub-tasks. The controller reads inputs from an input-tape, uses the data memory to store valuable signals and writes correct symbols to an output tape. The output symbols are a function of the outputs of each sub-network and the state of the data memory. Hence, the controller learns to load the weights of the appropriate program network to generate output symbols.

A wide range of experiments demonstrate that the Harvard Machine framework learns faster and performs better than the NTM and RNNs like LSTM, as the complexity of tasks increases.

In a multi-robot system, locating a team robot is an important issue. If robots

can refer to the location of team robots based on information through passive action

recognition without explicit communication, various advantages (e.g. improving security

for military purposes) can be obtained. Specifically, when team robots follow

the same motion rule based on information about adjacent robots, associations can

be found between robot actions. If the association can be analyzed, this can be a clue

to the remote robot. Using these clues, it is possible to infer remote robots which are

outside of the sensor range.

In this paper, a multi-robot system is constructed using a combination of Thymio

II robotic platforms and Raspberry pi controllers. Robots moving in chain-formation

take action using motion rules based on information obtained through passive action

recognition. To find associations between robots, a regression model is created using

Deep Neural Network (DNN) and Long Short-Term Memory (LSTM), one of state-of-art technologies.

The input data of the regression model is divided into historical data, which

are consecutive positions of the robot, and observed data, which is information about the

observed robot. Historical data is sequence data that is analyzed through the LSTM

layer. The accuracy of the regression model designed using DNN can vary depending

on the quantity and quality of the input. In this thesis, three different input situations

are assumed for comparison. First, the amount of observed data is different, second, the

type of observed data is different, and third, the history length is different. Comparative

models are constructed for each case, and prediction accuracy is compared to analyze

the effect of input data on the regression model. This exploration validates that these

methods from deep learning can reduce the communication demands in coordinated

motion of multi-robot systems

Recently, a well-designed and well-trained neural network can yield state-of-the-art results across many domains, including data mining, computer vision, and medical image analysis. But progress has been limited for tasks where labels are difficult or impossible to obtain. This reliance on exhaustive labeling is a critical limitation in the rapid deployment of neural networks. Besides, the current research scales poorly to a large number of unseen concepts and is passively spoon-fed with data and supervision.

To overcome the above data scarcity and generalization issues, in my dissertation, I first propose two unsupervised conventional machine learning algorithms, hyperbolic stochastic coding, and multi-resemble multi-target low-rank coding, to solve the incomplete data and missing label problem. I further introduce a deep multi-domain adaptation network to leverage the power of deep learning by transferring the rich knowledge from a large-amount labeled source dataset. I also invent a novel time-sequence dynamically hierarchical network that adaptively simplifies the network to cope with the scarce data.

To learn a large number of unseen concepts, lifelong machine learning enjoys many advantages, including abstracting knowledge from prior learning and using the experience to help future learning, regardless of how much data is currently available. Incorporating this capability and making it versatile, I propose deep multi-task weight consolidation to accumulate knowledge continuously and significantly reduce data requirements in a variety of domains. Inspired by the recent breakthroughs in automatically learning suitable neural network architectures (AutoML), I develop a nonexpansive AutoML framework to train an online model without the abundance of labeled data. This work automatically expands the network to increase model capability when necessary, then compresses the model to maintain the model efficiency.

In my current ongoing work, I propose an alternative method of supervised learning that does not require direct labels. This could utilize various supervision from an image/object as a target value for supervising the target tasks without labels, and it turns out to be surprisingly effective. The proposed method only requires few-shot labeled data to train, and can self-supervised learn the information it needs and generalize to datasets not seen during training.