Sexual abuse is a horrible reality for many children around the world. As technology improves the availability of encryption schemes and anonymity over the internet, the perpetrators of these acts are increasingly hard to track. There have been several advances in recent time to automate the work of trying to catch these perpetrators and especially image recognition has seen great promise. While image recognition is a natural approach to these subjects as many abuses are documented and shared between perpetrators, there are potentially many leads that go unexplored if only focusing on images and videos. This study evaluates how methods of supervised machine learning solely based on textual data can point us to posts on forums which are connected to the distribution of child sexual abusive material. Feature representation techniques such as word-vectors, paragraphvectors and the FastText algorithm were used in conjunction with supervised machine learning methods based on deep learning, including methods of multilayer perceptrons, convolutional neural networks and long-short term memory models. The models were trained and evaluated on a dataset based on forum posts from a Dark Net leak from last year, and are evaluated as well on text collected from websites that had been manually verified by Ecpat. Those models were compared to a baseline model based on logistic regression. It was found that those state-of-the-art models achieve a similar performance, all outperforming the 'benchmark' logistic regression model. Further improvements can be achieved based on the availability of more annotated data.

Neural networks based deep learning methods have gained significant success in several real world tasks: from machine translation to web recommendation, and it is also greatly improving the computer vision and the natural language processing. Compared with conventional machine learning techniques, neural network based deep learning do not require careful engineering and consideration domain expertise to design a feature extractor that transformed the raw data to a suitable internal representation. Its extreme efficacy on multiple levels of representation and feature learning ensures this type of approaches can process high dimensional data. It integrates the feature representation, learning and recognition into a systematical framework, which allows the learning starts at one level (i.e., being with raw input) and end at a higher slightly more abstract level. By simply stacking enough such transformations, very complex functions can be obtained. In general, high level feature representation facilitate the discrimination of patterns, and additionally can reduce the impact of irrelevant variations. However, previous studies indicate that deep composition of the networks make the training errors become vanished. To overcome this weakness, several techniques have been developed, for instance, dropout, stochastic gradient decent and residual network structures. In this study, we incorporates latent information into different network structures (e.g., restricted Boltzmann machine, recursive neural networks, and long short term memory). The conditional latent information reflects the high dimensional correlation existed in the data structure, and the typical network structure may not learn this kind of features due to limitation of the initial design (i.e., the network size the parameters). Similarly to residual nets, the conditional neural networks jointly learns the global features and local features, and the specifically designed network structure helps to incorporate the modulation derived from the probability distribution. The proposed models have been widely tested in different datasets, for instance, the conditional RBM has been applied to detect the speech components, and a language model based gated RBM has been used to recognize speech related EEG patterns. The conditional RNN has been tested in both general natural language modeling and medical notes prediction tasks. The results indicate that by introducing conditional branches in the conventional network structures, the latent features can be globally and locally learned.

Object Detection is widely used in many applications such as face detection, detecting vehicles and pedestrians on streets, and autonomous vehicles. Object detection not only includes recognizing and classifying objects in an image, but also localizes those objects and draws bounding boxes around them. Therefore, most of the successful object detection networks make use of neural network based image classifiers in conjunction with object detection techniques. Tensorflow Object Detection API, an open source framework based on Google's TensorFlow, allows us to create, train and deploy object detection models. This thesis mainly focuses on detecting objects kept in a refrigerator. To facilitate the object detection in a refrigerator, we have used Tensorflow Object Detection API to train and evaluate models such as SSD-MobileNet-v2, Faster R-CNN-ResNet-101, and R-FCN-ResNet-101. The models are tested as a) a pre-trained model and b) a fine-tuned model devised by fine-tuning the existing models with a training dataset for eight food classes extracted from the ImageNet database. The models are evaluated on a test dataset for the same eight classes derived from the ImageNet database to infer which works best for our application. The results suggest that the performance of Faster R-CNN is the best on the test food dataset with a mAP score of 81.74%, followed by R-FCN with a mAP of 80.33% and SSD with a mAP of 76.39%. However, the time taken by SSD for detection is considerably less than the other two models which makes it a viable option for our objective. The results provide substantial evidence that the SSD model is the most suitable model for deploying object detection on mobile devices with an accuracy of 76.39%. Our methodology and results could potentially help other researchers to design a custom object detector and further enhance the precision for their datasets. / Graduate

abstract: Discriminative learning when training and test data belong to different distributions is a challenging and complex task. Often times we have very few or no labeled data from the test or target distribution, but we may have plenty of labeled data from one or multiple related sources with different distributions. Due to its capability of migrating knowledge from related domains, transfer learning has shown to be effective for cross-domain learning problems. In this dissertation, I carry out research along this direction with a particular focus on designing efficient and effective algorithms for BioImaging and Bilingual applications. Specifically, I propose deep transfer learning algorithms which combine transfer learning and deep learning to improve image annotation performance. Firstly, I propose to generate the deep features for the Drosophila embryo images via pretrained deep models and build linear classifiers on top of the deep features. Secondly, I propose to fine-tune the pretrained model with a small amount of labeled images. The time complexity and performance of deep transfer learning methodologies are investigated. Promising results have demonstrated the knowledge transfer ability of proposed deep transfer algorithms. Moreover, I propose a novel Robust Principal Component Analysis (RPCA) approach to process the noisy images in advance. In addition, I also present a two-stage re-weighting framework for general domain adaptation problems. The distribution of source domain is mapped towards the target domain in the first stage, and an adaptive learning model is proposed in the second stage to incorporate label information from the target domain if it is available. Then the proposed model is applied to tackle cross lingual spam detection problem at LinkedIn’s website. Our experimental results on real data demonstrate the efficiency and effectiveness of the proposed algorithms. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2015

abstract: Computational visual aesthetics has recently become an active research area. Existing state-of-art methods formulate this as a binary classification task where a given image is predicted to be beautiful or not. In many applications such as image retrieval and enhancement, it is more important to rank images based on their aesthetic quality instead of binary-categorizing them. Furthermore, in such applications, it may be possible that all images belong to the same category. Hence determining the aesthetic ranking of the images is more appropriate. To this end, a novel problem of ranking images with respect to their aesthetic quality is formulated in this work. A new data-set of image pairs with relative labels is constructed by carefully selecting images from the popular AVA data-set. Unlike in aesthetics classification, there is no single threshold which would determine the ranking order of the images across the entire data-set. This problem is attempted using a deep neural network based approach that is trained on image pairs by incorporating principles from relative learning. Results show that such relative training procedure allows the network to rank the images with a higher accuracy than a state-of-art network trained on the same set of images using binary labels. Further analyzing the results show that training a model using the image pairs learnt better aesthetic features than training on same number of individual binary labelled images. Additionally, an attempt is made at enhancing the performance of the system by incorporating saliency related information. Given an image, humans might fixate their vision on particular parts of the image, which they might be subconsciously intrigued to. I therefore tried to utilize the saliency information both stand-alone as well as in combination with the global and local aesthetic features by performing two separate sets of experiments. In both the cases, a standard saliency model is chosen and the generated saliency maps are convoluted with the images prior to passing them to the network, thus giving higher importance to the salient regions as compared to the remaining. Thus generated saliency-images are either used independently or along with the global and the local features to train the network. Empirical results show that the saliency related aesthetic features might already be learnt by the network as a sub-set of the global features from automatic feature extraction, thus proving the redundancy of the additional saliency module. / Dissertation/Thesis / Masters Thesis Computer Science 2016

abstract: High-level inference tasks in video applications such as recognition, video retrieval, and zero-shot classification have become an active research area in recent years. One fundamental requirement for such applications is to extract high-quality features that maintain high-level information in the videos. Many video feature extraction algorithms have been purposed, such as STIP, HOG3D, and Dense Trajectories. These algorithms are often referred to as “handcrafted” features as they were deliberately designed based on some reasonable considerations. However, these algorithms may fail when dealing with high-level tasks or complex scene videos. Due to the success of using deep convolution neural networks (CNNs) to extract global representations for static images, researchers have been using similar techniques to tackle video contents. Typical techniques first extract spatial features by processing raw images using deep convolution architectures designed for static image classifications. Then simple average, concatenation or classifier-based fusion/pooling methods are applied to the extracted features. I argue that features extracted in such ways do not acquire enough representative information since videos, unlike images, should be characterized as a temporal sequence of semantically coherent visual contents and thus need to be represented in a manner considering both semantic and spatio-temporal information. In this thesis, I propose a novel architecture to learn semantic spatio-temporal embedding for videos to support high-level video analysis. The proposed method encodes video spatial and temporal information separately by employing a deep architecture consisting of two channels of convolutional neural networks (capturing appearance and local motion) followed by their corresponding Fully Connected Gated Recurrent Unit (FC-GRU) encoders for capturing longer-term temporal structure of the CNN features. The resultant spatio-temporal representation (a vector) is used to learn a mapping via a Fully Connected Multilayer Perceptron (FC-MLP) to the word2vec semantic embedding space, leading to a semantic interpretation of the video vector that supports high-level analysis. I evaluate the usefulness and effectiveness of this new video representation by conducting experiments on action recognition, zero-shot video classification, and semantic video retrieval (word-to-video) retrieval, using the UCF101 action recognition dataset. / Dissertation/Thesis / Masters Thesis Computer Science 2016

abstract: The widespread adoption of computer vision models is often constrained by the issue of domain mismatch. Models that are trained with data belonging to one distribution, perform poorly when tested with data from a different distribution. Variations in vision based data can be attributed to the following reasons, viz., differences in image quality (resolution, brightness, occlusion and color), changes in camera perspective, dissimilar backgrounds and an inherent diversity of the samples themselves. Machine learning techniques like transfer learning are employed to adapt computational models across distributions. Domain adaptation is a special case of transfer learning, where knowledge from a source domain is transferred to a target domain in the form of learned models and efficient feature representations. The dissertation outlines novel domain adaptation approaches across different feature spaces; (i) a linear Support Vector Machine model for domain alignment; (ii) a nonlinear kernel based approach that embeds domain-aligned data for enhanced classification; (iii) a hierarchical model implemented using deep learning, that estimates domain-aligned hash values for the source and target data, and (iv) a proposal for a feature selection technique to reduce cross-domain disparity. These adaptation procedures are tested and validated across a range of computer vision applications like object classification, facial expression recognition, digit recognition, and activity recognition. The dissertation also provides a unique perspective of domain adaptation literature from the point-of-view of linear, nonlinear and hierarchical feature spaces. The dissertation concludes with a discussion on the future directions for research that highlight the role of domain adaptation in an era of rapid advancements in artificial intelligence. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2017

abstract: To ensure system integrity, robots need to proactively avoid any unwanted physical perturbation that may cause damage to the underlying hardware. In this thesis work, we investigate a machine learning approach that allows robots to anticipate impending physical perturbations from perceptual cues. In contrast to other approaches that require knowledge about sources of perturbation to be encoded before deployment, our method is based on experiential learning. Robots learn to associate visual cues with subsequent physical perturbations and contacts. In turn, these extracted visual cues are then used to predict potential future perturbations acting on the robot. To this end, we introduce a novel deep network architecture which combines multiple sub- networks for dealing with robot dynamics and perceptual input from the environment. We present a self-supervised approach for training the system that does not require any labeling of training data. Extensive experiments in a human-robot interaction task show that a robot can learn to predict physical contact by a human interaction partner without any prior information or labeling. Furthermore, the network is able to successfully predict physical contact from either depth stream input or traditional video input or using both modalities as input. / Dissertation/Thesis / Masters Thesis Computer Science 2017

abstract: Alzheimer’s Disease (AD), a neurodegenerative disease is a progressive disease that affects the brain gradually with time and worsens. Reliable and early diagnosis of AD and its prodromal stages (i.e. Mild Cognitive Impairment(MCI)) is essential. Fluorodeoxyglucose (FDG) positron emission tomography (PET) measures the decline in the regional cerebral metabolic rate for glucose, offering a reliable metabolic biomarker even on presymptomatic AD patients. PET scans provide functional information that is unique and unavailable using other types of imaging. The computational efficacy of FDG-PET data alone, for the classification of various Alzheimer’s Diagnostic categories (AD, MCI (LMCI, EMCI), Control) has not been studied. This serves as motivation to correctly classify the various diagnostic categories using FDG-PET data. Deep learning has recently been applied to the analysis of structural and functional brain imaging data. This thesis is an introduction to a deep learning based classification technique using neural networks with dimensionality reduction techniques to classify the different stages of AD based on FDG-PET image analysis. This thesis develops a classification method to investigate the performance of FDG-PET as an effective biomarker for Alzheimer's clinical group classification. This involves dimensionality reduction using Probabilistic Principal Component Analysis on max-pooled data and mean-pooled data, followed by a Multilayer Feed Forward Neural Network which performs binary classification. Max pooled features result into better classification performance compared to results on mean pooled features. Additionally, experiments are done to investigate if the addition of important demographic features such as Functional Activities Questionnaire(FAQ), gene information helps improve performance. Classification results indicate that our designed classifiers achieve competitive results, and better with the additional of demographic features. / Dissertation/Thesis / Masters Thesis Computer Science 2017

abstract: Commercial buildings in the United States account for 19% of the total energy consumption annually. Commercial Building Energy Consumption Survey (CBECS), which serves as the benchmark for all the commercial buildings provides critical input for EnergyStar models. Smart energy management technologies, sensors, innovative demand response programs, and updated versions of certification programs elevate the opportunity to mitigate energy-related problems (blackouts and overproduction) and guides energy managers to optimize the consumption characteristics. With increasing advancements in technologies relying on the ‘Big Data,' codes and certification programs such as the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE), and the Leadership in Energy and Environmental Design (LEED) evaluates during the pre-construction phase. It is mostly carried out with the assumed quantitative and qualitative values calculated from energy models such as Energy Plus and E-quest. However, the energy consumption analysis through Knowledge Discovery in Databases (KDD) is not commonly used by energy managers to perform complete implementation, causing the need for better energy analytic framework. The dissertation utilizes Interval Data (ID) and establishes three different frameworks to identify electricity losses, predict electricity consumption and detect anomalies using data mining, deep learning, and mathematical models. The process of energy analytics integrates with the computational science and contributes to several objectives which are to 1. Develop a framework to identify both technical and non-technical losses using clustering and semi-supervised learning techniques. 2. Develop an integrated framework to predict electricity consumption using wavelet based data transformation model and deep learning algorithms. 3. Develop a framework to detect anomalies using ensemble empirical mode decomposition and isolation forest algorithms. With a thorough research background, the first phase details on performing data analytics on the demand-supply database to determine the potential energy loss reduction potentials. Data preprocessing and electricity prediction framework in the second phase integrates mathematical models and deep learning algorithms to accurately predict consumption. The third phase employs data decomposition model and data mining techniques to detect the anomalies of institutional buildings. / Dissertation/Thesis / Doctoral Dissertation Civil, Environmental and Sustainable Engineering 2017