Deep Learning

In this research, we propose to use deep learning to predict crimes in small neighborhoods (regions) of a city, by using historical crime data collected from the past. The motivation of crime predictions is that if we can predict the number crimes that will occur in a certain week then the city officials and law enforcement can prepare resources and manpower more effectively. Due to inherent connections between geographic regions and crime activities, the crime numbers in different regions (with respect to different time periods) are often correlated. Such correlation brings challenges and opportunities to employ deep learning to learn features from historical data for accurate prediction of the future crime numbers for each neighborhood. To leverage crime correlations between different regions, we convert crime data into a heat map, to show the intensity of crime numbers and the geographical distributions. After that, we design a deep learning framework to learn from such heat map for prediction.
In our study, we look at the crime reported in twenty different neighbourhoods in Vancouver, Canada over a twenty week period and predict the total crime count that will occur in the future. We will look at the number of crimes per week that have occurred in the span of ten weeks and predict the crime count for the following weeks.
The location of where the crimes occur is extracted from a database and plotted onto a heat map. The model we are using to predict the crime count consists of a CNN (Convolutional Neural Network) and a LSTM (Long-Short Term Memory) network attached to the CNN. The purpose of the CNN is to train the model spatially and understand where crimes occur in the images. The LSTM is used to train the model temporally and help us understand which week the crimes occur in time. By feeding the model heat map images of crime hot spots into the CNN and LSTM network, we will be able to predict the crime count and the most likely locations of the crimes for future weeks.

Machine learning has been utilized in bio-imaging in recent years, however as it is relatively new and evolving, some researchers who wish to utilize machine learning tools have limited access because of a lack of programming knowledge. In electron microscopy (EM), immunogold labeling is commonly used to identify the target proteins, however the manual annotation of the gold particles in the images is a time-consuming and laborious process. Conventional image processing tools could provide semi-automated annotation, but those require that users make manual adjustments for every step of the analysis. To create a new high-throughput image analysis tool for immuno-EM, I developed a deep learning pipeline that was designed to deliver a completely automated annotation of immunogold particles in EM images. The program was made accessible for users without prior programming experience and was also expanded to be used on different types of immuno-EM images.

The success of deep learning has renewed interest in applying neural networks and other machine learning techniques to most fields of data and signal processing, including communications. Advances in architecture and training lead us to consider new modem architectures that allow flexibility in design, continued learning in the field, and improved waveform coding. This dissertation examines neural network architectures and training methods suitable for demodulation in power-limited communication systems, such as those found in wireless sensor networks. Such networks will provide greater connection to the world around us and are expected to contain orders of magnitude more devices than cellular networks. A number of standard and proprietary protocols span this space, with modulations such as frequency-shift-keying (FSK), Gaussian FSK (GFSK), minimum shift keying (MSK), on-off-keying (OOK), and M-ary orthogonal modulation (M-orth). These modulations enable low-cost radio hardware with efficient nonlinear amplification in the transmitter and noncoherent demodulation in the receiver.

Engagement with educational instruction and related materials is an important part of learning and contributes to test performance. There are various measures of engagement including self-reports, observations, pupil diameter, and posture. With the challenges associated with obtaining accurate engagement levels, such as difficulties with measuring variations in engagement, the present study used a novel approach to predict engagement from posture by using deep learning. Deep learning was used to analyze a labeled outline of the participants and extract key points that are expected to predict engagement. In the first experiment two short lectures were presented and participants were tested on a lecture to motivate engagement. The next experiment had videos that varied in interest to understand whether a more interesting presentation engages participants more, therefore helping participants achieve higher comprehension scores. In a third experiment, one video was presented to attempt to use posture to predict comprehension rather than engagement. The fourth experiment had videos that varied in level of difficulty to determine whether a challenging topic versus an easier topic affects engagement. T-tests revealed that the more interesting Ted Talk was rated as more engaging, and for the fourth study, the more difficult video was rated as more engaging. Comparing average pupil sizes did not reveal significant differences that would relate to differences in the engagement scores, and average pupil dilation did not correlate with engagement. Analyzing posture through deep learning resulted in three accurate predictive models and a way to predict comprehension. Since engagement relates to learning, researchers and educators can benefit from accurate engagement measures.

Gliomas are an aggressive class of brain tumors that are associated with a better prognosis at a lower grade level. Effective differentiation and classification are imperative for early treatment. MRI scans are a popular medical imaging modality to detect and diagnosis brain tumors due to its capability to non-invasively highlight the tumor region. With the rise of deep learning, researchers have used convolution neural networks for classification purposes in this domain, specifically pre-trained networks to reduce computational costs. However, with various MRI modalities, MRI machines, and poor image scan quality cause different network structures to have different performance metrics. Each pre-trained network is designed with a different structure that allows robust results given specific problem conditions. This thesis aims to cover the gap in the literature to compare the performance of popular pre-trained networks on a controlled dataset that is different than the network trained domain.

Deep Neural Networks have been widely applied in many different applications and achieve significant improvement over classical machine learning techniques. However, training a neural network usually requires large amount of data, which is not guaranteed in some applications such as medical image classification. To address this issue, people propose to implement meta learning and ensemble learning techniques to make deep learning trainers more powerful. This thesis focuses on using deep learning equipped with meta learning and ensemble learning to study specific problems.
We first propose a new deep learning based method for suggestion mining.
The major challenges of suggestion mining include cross domain issue and the issues caused by unstructured and highly imbalanced data structure. To overcome these challenges, we propose to apply Random Multi-model Deep Learning (RMDL) which combines three different deep learning architectures (DNNs, RNNs and CNNs) and automatically selects the optimal hyper parameter to improve the robustness and flexibility of the model. Our experimental results on the SemEval-2019 competition Task 9 data sets demonstrate that our proposed RMDL outperforms most of the existing suggestion mining methods.

Deep learning models have been successfully applied to a variety of machine learning tasks, including image identification, image segmentation, object detection, speaker recognition, natural language processing, bioinformatics and drug discovery, among other things. This dissertation introduces Multi-Model Deep Learning (MMDL), a new ensemble deep learning approach for signal classification and event forecasting. The ultimate goal of the MMDL method is to improve classification and forecasting performances of individual classifiers by fusing results of participating deep learning models. The performance of such an ensemble model, however, depends heavily on the following two design features. Firstly, the diversity of the participating (or base) deep learning models is crucial. If all base deep learning models produce similar classification results, then combining these results will not provide much improvement. Thus, diversity is considered to be a key design feature of any successful MMDL system. Secondly, the selection of a fusion function, namely, a suitable function to integrate the results of all the base models, is important. In short, building an effective MMDL system is a complex and challenging process which requires deep knowledge of the problem context and a well-defined prediction process. The proposed MMDL method utilizes a bank of Convolutional Neural Networks (CNNs) and Stacked AutoEncoders (SAEs). To reduce the design complexity, a randomized generation process is applied to assign values to hyperparameters of base models. To speed up the training process, new feature extraction procedures which captures time-spatial characteristics of input signals are also explored. The effectiveness of the MMDL method is validated in this dissertation study with three real-world case studies. In the first case study, the MMDL model is applied to classify call types of groupers, an important fishery resource in the Caribbean that produces sounds associated with reproductive behaviors during yearly spawning aggregations. In the second case study, the MMDL model is applied to detect upcalls of North Atlantic Right Whales (NARWs), a type of endangered whales. NARWs use upcalls to communicate among themselves. In the third case study, the MMDL model is modified to predict seizure episodes. In all these cases, the proposed MMDL model outperforms existing state-of-the-art methods.

The success of deep learning in applications including computer vision, natural language processing, and even the game of Go can only be a orded by powerful computational resources and vast data sets. Data sets coming from the medical application are often much smaller and harder to acquire. Here a novel data approach is explained and used to demonstrate how to use deep learning as a step in data discovery, classi cation, and ultimately support for further investigation. Data sets used to illustrate these successes come from common ion-separation techniques that allow for gas samples to be quantitatively analyzed. The success of this data approach allows for the deployment of deep learning to smaller data sets.

Effective classification with imbalanced data is an important area of research, as high class imbalance is naturally inherent in many real-world applications, e.g. anomaly detection. Modeling such skewed data distributions is often very difficult, and non-standard methods are sometimes required to combat these negative effects. These challenges have been studied thoroughly using traditional machine learning algorithms, but very little empirical work exists in the area of deep learning with class imbalanced big data. Following an in-depth survey of deep learning methods for addressing class imbalance, we evaluate various methods for addressing imbalance on the task of detecting Medicare fraud, a big data problem characterized by extreme class imbalance. Case studies herein demonstrate the impact of class imbalance on neural networks, evaluate the efficacy of data-level and algorithm-level methods, and achieve state-of-the-art results on the given Medicare data set. Results indicate that combining under-sampling and over-sampling maximizes both performance and efficiency.

Smartphones and mobile tablets are rapidly growing, and very important nowadays. The most popular mobile operating system since 2012 has been Android. Android is an open source platform that allows developers to take full advantage of both the operating system and the applications itself. However, due to the open source community of an Android platform, some Android developers took advantage of this and created countless malicious applications such as Trojan, Malware, and Ransomware. All which are currently hidden in a large number of benign apps in official Android markets, such as Google PlayStore, and Amazon. Ransomware is a malware that once infected the victim’s device. It will encrypt files, unlock device system, and display a popup message which asks the victim to pay ransom in order to unlock their device or system which may include medical devices that connect through the internet. In this research, we propose to combine permission and API calls, then use Deep Learning techniques to detect ransomware apps from the Android market. Permissions setting and API calls are extracted from each app file by using a python library called AndroGuard. We are using Permissions and API call features to characterize each application, which can identify which application has potential to be ransomware or is benign. We implement our Android Ransomware Detection framework based on Keras, which uses MLP with back-propagation and a supervised algorithm. We used our method with experiments based on real-world applications with over 2000 benign applications and 1000 ransomware applications. The dataset came from ARGUS’s lab [1] which validated algorithm performance and selected the best architecture for the multi-layer perceptron (MLP) by trained our dataset with 6 various of MLP structures. Our experiments and validations show that the MLPs have over 3 hidden layers with medium sized of neurons achieved good results on both accuracy and AUC score of 98%. The worst score is approximately 45% to 60% and are from MLPs that have 2 hidden layers with large number of neurons.