Search Results (1 - 19 of 19 Results)

Sort By  
Sort Dir
 
Results per page  

Howard, Shaun MichaelDeep Learning for Sensor Fusion
Master of Sciences (Engineering), Case Western Reserve University, 2017, EECS - Computer and Information Sciences
The use of multiple sensors in modern day vehicular applications is necessary to provide a complete outlook of surroundings for advanced driver assistance systems (ADAS) and automated driving. The fusion of these sensors provides increased certainty in the recognition, localization and prediction of surroundings. A deep learning-based sensor fusion system is proposed to fuse two independent, multi-modal sensor sources. This system is shown to successfully learn the complex capabilities of an existing state-of-the-art sensor fusion system and generalize well to new sensor fusion datasets. It has high precision and recall with minimal confusion after training on several million examples of labeled multi-modal sensor data. It is robust, has a sustainable training time, and has real-time response capabilities on a deep learning PC with a single NVIDIA GeForce GTX 980Ti graphical processing unit (GPU).

Committee:

Wyatt Newman, Dr (Committee Chair); M. Cenk Cavusoglu, Dr (Committee Member); Michael Lewicki, Dr (Committee Member)

Subjects:

Artificial Intelligence; Computer Science

Keywords:

deep learning; sensor fusion; deep neural networks; advanced driver assistance systems; automated driving; multi-stream neural networks; feedforward; multilayer perceptron; recurrent; gated recurrent unit; long-short term memory; camera; radar;

Chen, ZhiangDeep-learning Approaches to Object Recognition from 3D Data
Master of Sciences, Case Western Reserve University, 2017, EMC - Mechanical Engineering
This thesis focuses on deep-learning approaches to recognition and pose estimation of graspable objects using depth information. Recognition and orientation detection from depth-only data is encoded by a carefully designed 2D descriptor from 3D point clouds. Deep-learning approaches are explored from two main directions: supervised learning and semi-supervised learning. The disadvantages of supervised learning approaches drive the exploration of unsupervised pretraining. By learning good representations embedded in early layers, subsequent layers can be trained faster and with better performance. An understanding of learning processes from a probabilistic perspective is concluded, and it paves the way for developing networks based on Bayesian models, including Variational Auto-Encoders. Exploitation of knowledge transfer--re-using parameters learned from alternative training data--is shown to be effective in the present application.

Committee:

Wyatt Newman, PhD (Advisor); M. Cenk Çavusoglu, PhD (Committee Member); Roger Quinn, PhD (Committee Member)

Subjects:

Computer Science; Medical Imaging; Nanoscience; Robotics

Keywords:

deep learning; 3D object recognition; semi-supervised learning; knowledge transfer

Bettaieb, Luc AlexandreA Deep Learning Approach To Coarse Robot Localization
Master of Sciences (Engineering), Case Western Reserve University, 2017, EECS - Electrical Engineering
This thesis explores the use of deep learning for robot localization with applications in re-localizing a mislocalized robot. Seed values for a localization algorithm are assigned based on the interpretation of images. A deep neural network was trained on images acquired in and associated with named regions. In application, the neural net was used to recognize a region based on camera input. By recognizing regions from the camera, the robot can be localized grossly, and subsequently refined with existing techniques. Explorations into different deep neural network topologies and solver types are discussed. A process for gathering training data, training the classifier, and deployment through a robot operating system (ROS) package is provided.

Committee:

Wyatt Newman (Advisor); Murat Cavusoglu (Committee Member); Gregory Lee (Committee Member)

Subjects:

Computer Science; Electrical Engineering; Robotics

Keywords:

robotics; localization; deep learning; neural networks; machine learning; state estimation; robots; robot; robot operating system; ROS; AMCL; monte carlo localization; particle filter; ConvNets; convolutional neural networks

Cui, ChenConvolutional Polynomial Neural Network for Improved Face Recognition
Doctor of Philosophy (Ph.D.), University of Dayton, 2017, Electrical and Computer Engineering
Deep learning is the state-of-art technology in pattern recognition, especially in face recognition. The robustness of the deep network leads a better performance when the size of the training set becomes larger and larger. Convolutional Neural Network (CNN) is one of the most popular deep learning technologies in the modern world. It helps obtain various features from multiple filters in the convolutional layer and performs well in the hand written digits classification. Unlike the unique structure of each hand written digit, face features are more complex, and many difficulties are existed for face recognition in current research field, such as the variations of lighting conditions, poses, ages, etc. So the limitation of the nonlinear feature fitting of the regular CNN appears in the face recognition application. In order to create a better fitting curve for face features, we introduce a polynomial structure to the regular CNN to increase the non-linearity of the obtained features. The modified architecture is named as Convolutional Polynomial Neural Network (CPNN). CPNN creates a polynomial input for each convolutional layer and captures the nonlinear features for better classification. We firstly prove the proposed concept with MNIST handwritten database and compare the proposed CPNN with regular CNN. Then, different parameters in CPNN are tested by CMU AMP face recognition database. After that, the performance of the proposed CPNN is evaluated on three different face databases: CMU AMP, Yale and JAFFE as well as the images captured in real world environment. The proposed CPNN obtains the best recognition rates (CMU AMP: 99.95%, Yale: 90.89%, JAFFE: 98.33%, Real World: 97.22%) when compared to other different machine learning technologies. We are planning to apply the state-of-art structures, such as inception and residual, to the current CPNN to increase the depth and stability as our future research work.

Committee:

Vijayan Asari (Advisor)

Subjects:

Artificial Intelligence; Bioinformatics; Computer Engineering; Electrical Engineering

Keywords:

Deep Learning, Convolutional Polynomial Neural Network, Face Recognition, Computer Vision, Image Processing

Jin, WenjingModeling of Machine Life Using Accelerated Prognostics and Health Management (APHM) and Enhanced Deep Learning Methodology
PhD, University of Cincinnati, 2016, Engineering and Applied Science: Mechanical Engineering
Machine health monitoring has advanced significantly for improving machine uptime and efficiency by providing proper fault detection and remaining useful life (RUL) prediction information to machine users. Despite these advancements, conventional condition monitoring (CM) techniques face several challenges in machine prognostics, including the ineffective RUL prediction modeling for machine under dynamic working regimes, and the lack of complete lifecycle data for modeling and validation, among others. To address these issues, this research introduces Accelerated Degradation Tests (ADT) with a deep learning technique, which is a novel method to improve machine life prediction accuracy under different working regimes for Prognostics and Health Management applications. This dissertation work highlights the mathematical framework of deep learning based machine life modeling under an ADT environment, including Constant Stress Accelerated Degradation Testing (CSADT) and Step-Stress ADT (SSADT) conditions. Since most CM features show no trend or indication of failure until a machine is approaching the end of its life, current RUL prediction techniques are not applicable in that they are only effective when incipient degradation is detected. This dissertation work applies feature enhancement to condition-based features using the enhanced Restricted Boltzmann Machine (RBM) method with a prognosability regularization term; afterwards, a similarity-based method is applied to predict machine life with the enhanced RBM features. In addition, this research has added varying stress conditions during experiments to replicate dynamic operation regimes. The stress variable, a type of regime variables, is input into Mixed-Variate RBM (MV-RBM) model. Therefore, a Regime Matrix based RBM (RM-RBM) is proposed to improve the feature prognosability and reduce the impact that the working stresses have on the features. Then the RBM features can be fused into a single health value which reflects the machine degradation. Finally, the developed machine “life-stress-degradation” model can effectively estimate the machine life under any given stress. The feasibility study is demonstrated through three groups of rotary machinery components run-to-failure tests datasets. The first two case studies focus on CSADT from two bearing run-to-failure test-beds. The first in-house bearing test demonstrates the effectiveness of applying an enhanced RBM with a prognosability regularization term to improve predictability of both the features and health value. The second bearing test dataset utilizes a similarity-based method to benchmark the RUL prediction results of the RBM features with other feature extraction methods. The third case study focuses on the issue of dynamic operating regimes; it is validated through a step-stress accelerated degradation test on a ball screw system. By integrating both the RM-RBM model and SSADT model in the PHM analysis framework, an innovative condition-based life-stress model for the linear motion system will be demonstrated.

Committee:

Jay Lee, Ph.D. (Committee Chair); Linxia Liao, Ph.D. (Committee Member); Teik Lim, Ph.D. (Committee Member); David Thompson, Ph.D. (Committee Member)

Subjects:

Mechanical Engineering; Mechanics

Keywords:

Prognostics and Health Management;Accelerated Life Testing;Deep Learning;Restricted Boltzmann Machine;RUL Prediction;Condition Monitoring

Adams, William A.Analysis of Robustness in Lane Detection using Machine Learning Models
Master of Science (MS), Ohio University, 2015, Electrical Engineering (Engineering and Technology)
An appropriate approach to incorporating robustness into lane detection algorithms is beneficial to autonomous vehicle applications and other problems relying on fusion methods. While traditionally rigorous empirical methods were developed for mitigating lane detection error, an evidence-based model-driven approach yields robust results using multispectral video as input to various machine learning models. Branching beyond the few network structures considered for image understanding applications, deep networks with unique optimization functions are demonstrably more robust while making fewer assumptions. This work adopts a simple framework for data collection; retrieving image patches for comparison via regression through a learning model. Along a horizontal scanline, the most probable sample is selected to retrain the network. Models include simple regressors, various autoencoders, and a few specialized deep networks. Samples are compared by robustness and the results favor deep and highly specialized network structures.

Committee:

Mehmet Celenk (Advisor); Jeffrey Dill (Committee Member); Maarten Uijt de Haag (Committee Member); Rida Benhaddou (Committee Member)

Subjects:

Artificial Intelligence; Automotive Engineering; Computer Science; Engineering

Keywords:

Machine Learning; ADAS; Lane Detection; Autoencoder; Regressor; Deep Network; Deep Learning

Nair, Binu MuraleedharanLearning Latent Temporal Manifolds for Recognition and Prediction of Multiple Actions in Streaming Videos using Deep Networks
Doctor of Philosophy (Ph.D.), University of Dayton, 2015, Electrical Engineering
Recognizing multiple types of actions appearing in a continuous temporal order from a streaming video is the key to many possible applications ranging from real-time surveillance to egocentric motion for human computer interaction. Current state of the art algorithms are more focused either on holistic video representation or on finding a specific activity in video sequences. But the major drawback is that these algorithms work only on applications pertaining to unconstrained video search from the web and requires the complete sequence for reporting what kind of actions are present. In this dissertation, we propose an algorithm to detect and recognize multiple actions in a streaming sequence at every instant. This approach was successful in recognizing the type of action being performed and also provides a percentage of completion of that action at every instant in real-time. This system is invariant to the number of frames and the speed at which the action is being performed. Apart from these benefits, the proposed model can also predict the motion descriptors at future instances corresponding to the action present. Since human motion is inherently continuous in nature, the algorithm presented in this dissertation computes novel motion descriptors based on the dense optical flow at every instant and evaluates their variations along the temporal domain using deep learning techniques. For each action type, we compute a non-linear transformation from motion descriptor space into the latent temporal space using stacked autoencoders where this transformation is learned from its training patterns. The latent features thus obtained, forms a temporal manifold where the transitions along it are modeled using the Conditional Restricted Boltzmann Machines (CRBMs). Using these trained autoencoders and CRBMs for every action type, we can make an inference into multiple latent temporal action manifolds at an instant from a set of streaming input frames. Our model achieved a high accuracy of 93% in recognizing actions per frame and was able to predict the future action instances with an accuracy of 84% for KTH dataset. Similarly, it was also tested with the UCF Sports dataset and achieved an accuracy of 84% in recognizing the action per-frame and around 69% of predictive capability. Therefore we believe that the proposed model can benefit applications in human computer interaction, gaming and IP surveillance where the action classification using temporal manifolds and its predictive capability are crucial.

Committee:

Kimberly Kendricks (Committee Member); Keigo Hirakawa (Committee Member); Raul Ordonez (Committee Member); Vijayan Asari (Committee Chair)

Subjects:

Computer Engineering; Computer Science; Electrical Engineering; Statistics

Keywords:

motion descriptors, shape descriptors, principal component analysis, neural networks, autoencoder, restricted Boltzmann machine, conditional restricted Boltzmann machine, deep learning, action recognition, latent temporal manifold, action localization

Wang, PengSTOCHASTIC MODELING AND UNCERTAINTY EVALUATION FOR PERFORMANCE PROGNOSIS IN DYNAMICAL SYSTEMS
Doctor of Philosophy, Case Western Reserve University, 2017, EMC - Mechanical Engineering
The objective of this research is to advance the state of predictive science by integrating physics-based stochastic modeling methods with data analytic techniques to improve the accuracy and reliability of probabilistic prediction of event occurrence in dynamical systems, such as manufacturing machines or processes. To accomplish the above described objective, a hybrid modeling method is developed in this research, which integrates physical models for system performance degradation with stochastic modeling (realized by Bayesian inference) and data analytics (e.g. deep learning), to enable non-linear and non-Gaussian system modeling. The research addresses four fundamental questions : 1) how to quantify the accuracy and confidence in system tracking and performance variation prediction, given the limited observability; 2) how to incorporate uncertainty arising from varying operating conditions and environmental disturbance into prognostic models, 3) how to effectively fuse different data analytic techniques into one prognostic model, in order to take advantage of the strength of each of the techniques; and 4) how to improve the computational efficiency of the modeling and prediction process by leveraging emerging infrastructure enabled by cloud computing, for on-line, real-time, and remote prognosis. Specific research tasks include: 1) deriving system health indicators as a function of operating conditions measured by sensors through machine learning techniques (e.g. deep leaning); 2) developing stochastic models of system degradation based on multi-mode particle filter with adaptive resampling capability, to track variations in health indicators for prediction of performance deterioration with time-varying degradation rates and/or modes; 3) developing an optimization method to detect transient changes in health indicators due to abrupt fault occurrence; and 4) improving computational efficiency of particle filtering by modifying the sampling and resampling strategies and leveraging cloud-based computing. This research contributes to the fundamental theories of state estimation, tracking, prediction, and uncertainty evaluation. These theories can provide guidance to decision-making in maintenance policy of a wide range of dynamical systems, e.g., RUL prediction for aircraft engines for improved spare part management, fault diagnosis of HVAC systems in residential and commercial buildings, tool wear prediction in machine tools for intelligent manufacturing, etc.

Committee:

Robert Gao (Advisor); Roger Quinn (Committee Member); Bo Li (Committee Member); Wojbor Woyczynski (Committee Member)

Subjects:

Mechanical Engineering

Keywords:

Prognosis; Bayesian inference; Particle filtering; Deep learning; Dynamical systems

Chen, HuaFPGA Based Multi-core Architectures for Deep Learning Networks
Master of Science (M.S.), University of Dayton, 2015, Electrical Engineering
Deep learning a large scalable network architecture based on neural network. It is currently an extremely active research area in machine learning and pattern recognition society. They have diverse uses including pattern recognition, signal processing, image processing, image compression, classification of remote sensing data, and big data processing. Interest in specialized architectures for accelerating deep learning networks has increased significantly because of their ability to reduce power, increase performance, and allow fault tolerant computing. Specialized neuromorphic architectures could provide high performance at extreme low powers for these applications. This thesis concentrates on the implementation of multi-core neuromorphic network architecture on FPGA. Hardware prototyping of wormhole router unit is developed to control transmission of data packets running through between cores. Router units connect multiple cores into a large scalable network. This network is programmed on a Stratix IV FPGA board. Additionally, a memory initialization system is design inside the core to realize external network configuration. In this approaching, different applications could be mapped on the network without repeating FPGA compilation. One application called Image Edge Detection is mapped on the network. Finally this network outputs the desired image and demonstrate 3.4x run time efficiency and 3.6x energy-delay efficiency by FPGA implementation.

Committee:

Tarek Taha, Dr. (Advisor)

Subjects:

Electrical Engineering

Keywords:

Deep learning network; FPGA; neuromorphic processor; Wormhole router

Righi, Rebecca A.The Impact of Laptop Computers on Student Learning Behaviors as Perceived by Classroom Teachers
Master of Education, University of Toledo, 2012, Educational Administration and Supervision
The purpose of this study was to determine the impact of laptop computers on student learning behaviors. Each student and teacher was equipped with a laptop computer in which they had 24/7 access. Qualitative research methodology was used in this study and the data consisted of classroom observations, a review of the teachers’ lesson plans, and in-depth interviews with five classroom teachers. The results of this study revealed that laptop computers had a positive impact on student learning behaviors. Students were engaged in the learning process, produced higher quality work, and had improved communication with their teachers when they had access to laptop computers. Through analysis of the data, the researcher suggested that the changes in student behavior occurred because of personalized learning for each student, access to multiple materials and media, and the laptop computer serving as assistive technology.

Committee:

Cynthia Beekley, PhD (Committee Chair); Nancy Staub, PhD (Committee Member); Deb Gentry, PhD (Committee Member)

Subjects:

Education; Educational Technology

Keywords:

one-to-one program; laptop computers; learning behaviors; student engagement; deep learning

Putchala, Manoj KumarDeep Learning Approach for Intrusion Detection System (IDS) in the Internet of Things (IoT) Network using Gated Recurrent Neural Networks (GRU)
Master of Science (MS), Wright State University, 2017, Computer Science
The Internet of Things (IoT) is a complex paradigm where billions of devices are connected to a network. These connected devices form an intelligent system of systems that share the data without human-to-computer or human-to-human interaction. These systems extract meaningful data that can transform human lives, businesses, and the world in significant ways. However, the reality of IoT is prone to countless cyber-attacks in the extremely hostile environment like the internet. The recent hack of 2014 Jeep Cherokee, iStan pacemaker, and a German steel plant are a few notable security breaches. To secure an IoT system, the traditional high-end security solutions are not suitable, as IoT devices are of low storage capacity and less processing power. Moreover, the IoT devices are connected for longer time periods without human intervention. This raises a need to develop smart security solutions which are light-weight, distributed and have a high longevity of service. Rather than per-device security for numerous IoT devices, it is more feasible to implement security solutions for network data. The artificial intelligence theories like Machine Learning and Deep Learning have already proven their significance when dealing with heterogeneous data of various sizes. To substantiate this, in this research, we have applied concepts of Deep Learning and Transmission Control Protocol/Internet Protocol (TCP/IP) to build a light-weight distributed security solution with high durability for IoT network security. First, we have examined the ways of improving IoT architecture and proposed a light-weight and multi-layered design for an IoT network. Second, we have analyzed the existingapplications of Machine Learning and Deep Learning to the IoT and Cyber-Security. Third, we have evaluated deep learning’s Gated Recurrent Neural Networks (LSTM and GRU) on the DARPA/KDD Cup '99 intrusion detection data set for each layer in the designed architecture. Finally, from the evaluated metrics, we have proposed the best neural network design suitable for the IoT Intrusion Detection System. With an accuracy of 98.91% and False Alarm Rate of 0.76 %, this unique research outperformed the performance results of existing methods over the KDD Cup ’99 dataset. For this first time in the IoT research, the concepts of Gated Recurrent Neural Networks are applied for the IoT security.

Committee:

Michelle Cheatham, Ph.D. (Advisor); Adam Bryant, Ph.D. (Committee Member); Mateen Rizki, Ph.D. (Committee Member)

Subjects:

Computer Science

Keywords:

Deep Learning; Internet of Things; Machine Learning; Gated Recurrent Unit; Recurrent Neural Networks

McCoppin, Ryan R.An Evolutionary Approximation to Contrastive Divergence in Convolutional Restricted Boltzmann Machines
Master of Science (MS), Wright State University, 2014, Computer Science
Deep learning is an emerging area in machine learning that exploits multi-layered neural networks to extract invariant relationships from large data sets. Deep learning uses layers of non-linear transformations to represent data in abstract and discrete forms. Several different architectures have been developed over the past few years specifically to process images including the Convolutional Restricted Boltzmann Machine. The Boltzmann Machine is trained using contrastive divergence, a depth-first gradient based training algorithm. Gradient based training methods have no guarantee of reaching an optimal solution and tend to search a limited region of the solution space. In this thesis, we present an alternative method for synthesizing deep networks using evolutionary algorithms. This is a breadth-first stochastic search process that utilizes reconstruction error along with additional properties to encourage evolution of unique features. Using this technique, potentially a larger region of the solution space is explored allowing identification of different types of solutions using less training data. The process of developing this method is discussed along with its potential as a viable replacement to contrastive divergence.

Committee:

Mateen Rizki, Ph.D. (Advisor); Micheal Raymer, Ph.D. (Committee Member); John Gallagher, Ph.D. (Committee Member)

Subjects:

Computer Science

Keywords:

convolutional restricted boltzmann machine; CRBM; evolutionary algorithm; contrastive divergence; RBM; machine learning; deep learning

Liu, MenghanPULMONARY FUNCTION MONITORING USING PORTABLE ULTRASONOGRAPHY AND PRIVACY-PRESERVING LEARNING
Master of Sciences, Case Western Reserve University, 2017, EECS - Computer Engineering
Personal health monitoring system in home environment has gained more and more attention. In the personal data transmission and analysis, privacy is an important con- cern. In this thesis, I present a privacy-preserving health monitoring architecture, which can extract respiratory signs from ultrasound images and collaboratively build deep learning model for classifying health status. The architecture contains a global server and several local sites. Each local site consists of an ultrasound probe and a tablet. Per- formance of the system is evaluated with several experiments. The error of respiratory rate measurement is less than 0.5 time/minute, and the average error of tidal volume estimation is about 0.1 L. Performance of privacy-preserving deep learning architecture is tested using a human activity recognition dataset. The reconstructed rate could keep 90% in different scenarios. In conclusion, the proposed monitoring system is feasible for personal health monitoring.

Committee:

Huang Ming-Chun (Committee Chair); Danel Saab, G. (Committee Member); Soumyajit Mandal (Committee Member)

Subjects:

Computer Engineering; Health Care

Keywords:

Ultrasonography, Personal healthcare, Pulmonary function, Privacy-preserving deep learning

Han, KunSupervised Speech Separation And Processing
Doctor of Philosophy, The Ohio State University, 2014, Computer Science and Engineering
In real-world environments, speech often occurs simultaneously with acoustic interference, such as background noise or reverberation. The interference usually leads to adverse effects on speech perception, and results in performance degradation in many speech applications, including automatic speech recognition and speaker identification. Monaural speech separation and processing aim to separate or analyze speech from interference based on only one recording. Although significant progress has been made on this problem, it is a widely regarded challenge. Unlike traditional signal processing, this dissertation addresses the speech separation and processing problems using machine learning techniques. We first propose a classification approach to estimate the ideal binary mask (IBM) which is considered as a main goal of sound separation in computational auditory scene analysis (CASA). We employ support vector machines (SVMs) to classify time-frequency (T-F) units as either target-dominant or interference-dominant. A rethresholding method is incorporated to improve classification results and maximize hit minus false alarm rates. Systematic evaluations show that the proposed approach produces accurate estimated IBMs. In a supervised learning framework, the issue of generalization to conditions different from those in training is very important. We then present methods that require only a small training corpus and can generalize to unseen conditions. The system utilizes SVMs to learn classification cues and then employs a rethresholding technique to estimate the IBM. A distribution fitting method is introduced to generalize to unseen signal-to-noise ratio conditions and voice activity detection based adaptation is used to generalize to unseen noise conditions. In addition, we propose to use a novel metric learning method to learn invariant speech features in the kernel space. The learned features encode speech-related information and can generalize to unseen noise conditions. Experiments show that the proposed approaches produce high quality IBM estimates under unseen conditions. Besides background noise, room reverberation is another major source of signal degradation in real environments. Reverberation when combined with background noise is particularly disruptive for speech perception and many applications. We perform dereverberation and denoising using supervised learning. A deep neural network (DNN) is trained to directly learn a spectral mapping from the spectrogram of corrupted speech to that of clean speech. The spectral mapping approach substantially attenuates the distortion caused by reverberation and background noise, leading to improvement of predicted speech intelligibility and quality scores, as well as speech recognition rates. Pitch is one of the most important characteristics of speech signals. Although pitch tracking has been studied for decades, it is still challenging to estimate pitch from speech in the presence of strong noise. We estimate pitch using supervised learning, where probabilistic pitch states are directly learned from noisy speech data. We investigate two alternative neural networks modeling pitch state distribution given observations, i.e., a feedforward DNN and a recurrent deep neural network (RNN). Both DNNs and RNNs produce accurate probabilistic outputs of pitch states, which are then connected into pitch contours by Viterbi decoding. Experiments show that the proposed algorithms are robust to different noise conditions.

Committee:

DeLiang Wang (Advisor); Eric Fosler-Lussier (Committee Member); Mikhail Belkin (Committee Member)

Subjects:

Computer Science

Keywords:

Supervised learning; Speech separation; Speech processing; Machine learning; Deep Learning; Pitch estimation; Speech Dereverberation; Deep neural networks; Support vector machines

Liu, YeApplication of Convolutional Deep Belief Networks to Domain Adaptation
Master of Science, The Ohio State University, 2014, Computer Science and Engineering
Deep Belief Networks are instances of deep learning that have recently become prominent in machine learning. A recent extension considers adding convolution to the model, yielding better auto feature extraction for domains such as computer vision and natural language processing. The scope of this thesis is to explore the capability of Convolutional Deep Belief Networks (CDBN) in solving domain adaptation problem, for which other machine learning techniques have been provided having some level of both effectiveness and limitation. This thesis is part of a larger project which aims at solving the domain adaptation problems and producing higher accuracy in classification using deep learning algorithms. This project paves the way for further experiments in this direction by yielding promising results when applying CDBN to the domain adaptation problem. It shows that CDBNs are capable of capturing the meaningful features and filtering out irrelevant noise caused by the change of domains.

Committee:

Brian Kulis (Advisor); James Davis (Committee Member)

Subjects:

Computer Science

Keywords:

deep learning, Convolution, Convolutional Deep Belief Networks, features, image classification

Aspiras, Theus HerreraHierarchical Autoassociative Polynomial Network for Deep Learning of Complex Manifolds
Doctor of Philosophy (Ph.D.), University of Dayton, 2015, Electrical Engineering
Artificial neural networks are an area of research that has been explored extensively. With the formation of these networks, models of biological neural networks can be created mathematically for several different purposes. The neural network architecture being explored here is the nonlinear line attractor (NLA) network, which uses a polynomial weighting scheme instead of a linear weighting scheme for specific tasks. We have conducted research on this architecture and found that it works well to converge towards a specific trained pattern and diverge with untrained patterns. We have also improved the architecture with a Gaussian weighting scheme, which provides a modularity in the architecture and reduces redundancy in the network. Testing on the new weighting scheme improves network on different datasets gave better convergence characteristics, quicker training times, and improved recognition rates. The NLA architecture, however, is not able to reduce the dimensionality, thus a nonlinear dimensionality reduction technique is used. To improve the architecture further, we must be able to decompose the NLA architecture further to alleviate problems in the original structures and allow further improvements. We propose a hierarchical autoassociative polynomial network (HAP Net) which reorders the NLA architecture to include different ways to use polynomial weighting. In each layer, we can have orders of each input connected by a weight set, which can be trained by a backpropagation algorithm. By combining different architectures based on the understanding of MLP, attractor, and modular networks, we create a multi-purpose architecture including all aspects of the previous architecture which is far improved for classification and recognition tasks. Experiments conducted on the standard dataset, MNIST, shows very promising results of the HAP Net framework. Research work is progressing in evaluating performance on HAP Net on various datasets and also incorporating advanced learning strategies, convolutional neural networks, and extreme learning machine to investigate the performance.

Committee:

Vijayan Asari, Ph.D. (Committee Chair); Raul Ordonez, Ph.D (Committee Member); Eric Balster, Ph.D. (Committee Member); Wesam Sakla, Ph.D. (Committee Member)

Subjects:

Computer Engineering; Electrical Engineering

Keywords:

Polynomial Neural Network; Complex Manifolds; Deep Learning; Nonlinear Weighting; Modular; Classification; MNIST; HAP net

Farouni, TarekAn Overview of Probabilistic Latent Variable Models with an Application to the Deep Unsupervised Learning of Chromatin States
Doctor of Philosophy, The Ohio State University, 2017, Psychology
The following dissertation consists of two parts. The first part presents an overview of latent variable models from a probabilistic perspective. The main goal of the overview is to give a birds-eye view of the topographic structure of the space of latent variable models in light of recent developments in statistics and machine learning that show how seemingly unrelated models and methods are in fact intimately related to each other. In the second part of the dissertation, we apply a Deep Latent Gaussian Model (DLGM) to high-dimensional, high-throughput functional epigenomics datasets with the goal of learning a latent representation of functional regions of the genome, both across DNA sequence and across cell-types. In the latter half of the dissertation, we first demonstrate that the trained generative model is able to learn a compressed two-dimensional latent representation of the data. We then show how the learned latent space is able to capture the most salient patterns of dependencies in the observations such that synthetic samples simulated from the latent manifold are able to reconstruct the same patterns of dependencies we observe in data samples. Lastly, we provide a biological interpretation of the learned latent manifold in terms of a continuous histone code and explain both the relevance and significance of the proposed generative approach to the problem of identifying functional regulatory regions from epigenomic marks.

Committee:

Robert Cudeck (Advisor); Ewy Mathé (Committee Member); Zhong-Lin Lu (Committee Member); Paul DeBoeck (Committee Member)

Subjects:

Bioinformatics; Quantitative Psychology; Statistics

Keywords:

Probabilistic Latent Variable Models; Deep Generative Models; Deep Learning; Chromatin States; Histone Code; Epigenomics

Shakeel, AmlaanService robot for the visually impaired: Providing navigational assistance using Deep Learning
Master of Science, Miami University, 2017, Computational Science and Engineering
Assistive technology helps improve the day to day activities for people with disabilities. One of the methods utilized by assistive technologists employs the use of robots. These are called service robots. This thesis explores the idea of a service robot for the visually impaired to assist with navigation and is inspired by the use of guide dogs. The focus of this thesis is to develop a robot to achieve autonomous indoor navigation using computer vision to identify image based goals in an unfamiliar environment. The method presented in this thesis utilizes a deep learning framework, called Faster R-CNN, to train a computer to classify and localize exit signs in real time. A proof of concept is presented using NVIDIA Jetson, and TurtleBot, a robot kit, which runs a robot software development framework Robot Operating System (ROS). The model is trained successfully using Faster R-CNN and is validated. The model is used for real-time object classification on the prototype robot.

Committee:

Yamuna Rajasekhar (Advisor); John Femiani (Committee Member); Donald Ucci (Committee Member)

Subjects:

Computer Science; Electrical Engineering; Robotics

Keywords:

Assistive technology; Deep learning; Robotics; Indoor navigation; Computer vision; Robot Operating System; ROS; Caffe; Faster R-CNN; Convolutional Neural Networks; CNN; Microsoft Kinect; Service robots; visually impaired; mobility; depth perception

Ghayoumi, MehdiFACIAL EXPRESSION ANALYSIS USING DEEP LEARNING WITH PARTIAL INTEGRATION TO OTHER MODALITIES TO DETECT EMOTION
PHD, Kent State University, 2017, College of Arts and Sciences / Department of Computer Science
Analysis of human emotion is very important as the field of social robotics where a new generation of humanoids and other smart devices will interact with humans. Emotional expression is a universal language for interaction with humans. Understanding human emotions is a necessary and important step for human-computer interaction. Human emotion is expressed as a complex combination of facial expressions, speech (including silence) and gestures postures, various limb-motions, gaze, and blinking. Multiple research models have been developed for limited facial expression analysis, speech based emotion analysis, limited models for gesture analysis and their limited integration. However, such analysis is limited to single frame analysis time-efficiency, limited handling of occlusion, notion of colors in facial expression analysis, lack of exploitation of symmetry, lack of dynamic change in assigning weight between the modalities based upon environmental requirement and six basic emotions. This research develops a convolutional neural network based deep learning model that recognizes human facial expressions exploiting a combination of symmetrical representation to handle occlusion; a unified model based upon transforming facial muscle motion to geometric feature points; fusion of multiple modalities and fast hashing techniques for real-time emotion recognition. It also proposes a new model for recognition of mixed-emotion in real-time.

Committee:

Arvind K. Bansal (Advisor); Javed I. Khan (Committee Member); Cheng Chang Lu (Committee Member); Stephen B. Fountain (Committee Member); William E. Merriman (Committee Member)

Subjects:

Artificial Intelligence; Computer Science; Robotics; Robots

Keywords:

Human Computer Interaction, Emotion, Facial Expression, Deep Learning, Convolutional Neural Networks, Social Robots