Deep neural networks (DNNs) and deep learning approaches yield state-of-the-art performance in a range of machine learning tasks, including automatic speech recognition. The multi-layer transformations and activation functions in DNNs, or related network variations, allow complex and difficult data to be well modelled. However, the highly distributed representations associated with these models make it hard to interpret the parameters. The whole neural network is commonly treated a ``black box''. The behaviours of activation functions and the meanings of network parameters are rarely controlled in the standard DNN training. Though a sensible performance can be achieved, the lack of interpretations to network structures and parameters causes better regularisation and adaptation on DNN models challenging. In regularisation, parameters have to be regularised universally and indiscriminately. For instance, the widely used L2 regularisation encourages all parameters to be zeros. In adaptation, it requires to re-estimate a large number of independent parameters. Adaptation schemes in this framework cannot be effectively performed when there are limited adaptation data. This thesis investigates structured deep neural networks. Special structures are explicitly designed, and they are imposed with desired interpretation to improve DNN regularisation and adaptation. For regularisation, parameters can be separately regularised based on their functions. For adaptation, parameters can be adapted in groups or partially adapted according to their roles in the network topology. Three forms of structured DNNs are proposed in this thesis. The contributions of these models are presented as follows. The first contribution of this thesis is the multi-basis adaptive neural network. This form of structured DNN introduces a set of parallel sub-networks with restricted connections. The design of restricted connectivity allows different aspects of data to be explicitly learned. Sub-network outputs are then combined, and this combination module is used as the speaker-dependent structure that can be robustly estimated for adaptation. The second contribution of this thesis is the stimulated deep neural network. This form of structured DNN relates and smooths activation functions in regions of the network. It aids the visualisation and interpretation of DNN models but also has the potential to reduce over-fitting. Novel adaptation schemes can be performed on it, taking advantages of the smooth property that the stimulated DNN offer. The third contribution of this thesis is the deep activation mixture model. Also, this form of structured DNN encourages the outputs of activation functions to achieve a smooth surface. The output of one hidden layer is explicitly modelled as the sum of a mixture model and a residual model. The mixture model forms an activation contour, and the residual model depicts fluctuations around this contour. The smoothness yielded by a mixture model helps to regularise the overall model and allows novel adaptation schemes.

Autoregressive models factorize a multivariate joint probability distribution into a product of one-dimensional conditional distributions. The variables are assigned an ordering, and the conditional distribution of each variable modelled using all variables preceding it in that ordering as predictors. Calculating normalized probabilities and sampling has polynomial computational complexity under autoregressive models. Moreover, binary autoregressive models based on neural networks obtain statistical performances similar to that of some intractable models, like restricted Boltzmann machines, on several datasets. The use of autoregressive probability density estimators based on neural networks to model real-valued data, while proposed before, has never been properly investigated and reported. In this thesis we extend the formulation of neural autoregressive distribution estimators (NADE) to real-valued data; a model we call the real-valued neural autoregressive density estimator (RNADE). Its statistical performance on several datasets, including visual and auditory data, is reported and compared to that of other models. RNADE obtained higher test likelihoods than other tractable models, while retaining all the attractive computational properties of autoregressive models. However, autoregressive models are limited by the ordering of the variables inherent to their formulation. Marginalization and imputation tasks can only be solved analytically if the missing variables are at the end of the ordering. We present a new training technique that obtains a set of parameters that can be used for any ordering of the variables. By choosing a model with a convenient ordering of the dimensions at test time, it is possible to solve any marginalization and imputation tasks analytically. The same training procedure also makes it practical to train NADEs and RNADEs with several hidden layers. The resulting deep and tractable models display higher test likelihoods than the equivalent one-hidden-layer models for all the datasets tested. Ensembles of NADEs or RNADEs can be created inexpensively by combining models that share their parameters but differ in the ordering of the variables. These ensembles of autoregressive models obtain state-of-the-art statistical performances for several datasets. Finally, we demonstrate the application of RNADE to speech synthesis, and confirm that capturing the phone-conditional dependencies of acoustic features improves the quality of synthetic speech. Our model generates synthetic speech that was judged by naive listeners as being of higher quality than that generated by mixture density networks, which are considered a state-of-the-art synthesis technique.

abstract: Over the past decade, advancements in neural networks have been instrumental in achieving remarkable breakthroughs in the field of computer vision. One of the applications is in creating assistive technology to improve the lives of visually impaired people by making the world around them more accessible. A lot of research in convolutional neural networks has led to human-level performance in different vision tasks including image classification, object detection, instance segmentation, semantic segmentation, panoptic segmentation and scene text recognition. All the before mentioned tasks, individually or in combination, have been used to create assistive technologies to improve accessibility for the blind. This dissertation outlines various applications to improve accessibility and independence for visually impaired people during shopping by helping them identify products in retail stores. The dissertation includes the following contributions; (i) A dataset containing images of breakfast-cereal products and a classifier using a deep neural (ResNet) network; (ii) A dataset for training a text detection and scene-text recognition model; (iii) A model for text detection and scene-text recognition to identify product images using a user-controlled camera; (iv) A dataset of twenty thousand products with product information and related images that can be used to train and test a system designed to identify products. / Dissertation/Thesis / Masters Thesis Computer Science 2020

AN ABSTRACT OF THE DISSERTATION OFSherri Houmadi, for the Doctor of Philosophy degree in Engineering Science, presented on March 27, 2020, at Southern Illinois University Carbondale. TITLE: THE APPLICATION OF CONVOLUTIONAL NEURAL NETWORKS TO CLASSIFY PAINT DEFECTSMAJOR PROFESSOR: Dr. Julie DunstonDespite all of the technological advancements in computer vision, many companies still utilize human visual inspection to determine whether parts are good or bad. It is particularly challenging for humans to inspect parts in a fast-moving manufacturing environment. Such is the case at Aisin Manufacturing Illinois where this study will be testing the use of convolutional neural networks (CNNs) to classify paint defects on painted outside door handles and caps for automobiles. Widespread implementation of vision systems has resulted in advancements in machine learning. As the field of artificial intelligence (AI) evolves and improvement are made, diverse industries are adopting AI models for use in their applications. Medical imaging classification using neural networks has exploded in recent years. Convolutional neural networks have proven to scale very well for image classification models by extracting various features from the images. A goal of this study is to create a low-cost machine learning model that is able to quickly classify paint defects in order to identify rework parts that can be repaired and shipped. The central thesis of this doctoral work is to test a machine learning model that can classify the paint defects based on a very small dataset of images, where the images are taken with a smartphone camera in a manufacturing setting. The end goal is to train the model for an overall accuracy rate of at least 80%. By using transfer learning and balancing the class datasets, the model was trained to achieve an overall accuracy rate of 82%.

Surveillance cameras are a great support in crime investigation and proximity alarms and play a vital role in public safety. However current surveillance systems require continuous human supervision for monitoring. The primary goal of the thesis is to prevent firearm-related violence and injuries. Automatic firearm detection enhances security and safety among people. Therefore, introducing a Deep Learning Object Detection model to detect Firearms and alert the corresponding police department is the main motivation. Visual Object Detection is a fundamental recognition problem in computer vision that aims to find objects of certain target classes with precise localization of input image and assign it to the corresponding label. However, there are some challenges arising to the wide variations in shape, size, appearance, and occlusions by the weapon carrier. There are other objections to the selection of best object detection model. So, three deep learning models are selected, explained and shown the differences in detecting the firearms. The dataset in this thesis is the customized selection of different categories of firearm collection like the pistol, revolver, handgun, bullet, rifle along with human detection. The entire dataset is annotated manually in pascalvoc format. Date augmentation technique has been used to enlarge our dataset and facilitate in detecting firearms that re deformed and having occlusion properties.. To detect firearms this thesis developed and practiced unified networks like SSD and two-stage object detectors like faster RCNN. SSD is easy to understand and detect objects however it fails to detect smaller objects. Faster RCNN are efficient and able to detect smaller firearms in the scene. Each class has attained more than 90% of confidence score.

京都大学 / 新制・課程博士 / 博士(理学) / 甲第23361号 / 理博第4732号 / 新制||理||1679(附属図書館) / 京都大学大学院理学研究科物理学・宇宙物理学専攻 / (主査)教授 田中 貴浩, 准教授 久徳 浩太郎, 教授 萩野 浩一 / 学位規則第4条第1項該当 / Doctor of Science / Kyoto University / DFAM

Deep Neural Networks (DNNs) are the algorithm of choice for various applications that require modeling large datasets, such as image classification, object detection and natural language processing. DNNs present highly parallel workloads<br>that lead to the need of custom hardware accelerators. Deep Learning (DL) models specialized on different tasks require a programmable custom hardware, and a compiler to efficiently translate various DNNs into an efficient dataflow to be executed on the accelerator. Given a DNN oriented custom instructions set, various compilation phases are needed to generate efficient code and maintain generality to support<br>many models. Different compilation phases need to have different levels of hardware awareness so that it exploits the hardware’s full potential, while abiding with the hardware constraints. The goal of this work is to present a compiler workflow and its hardware aware optimization passes for a custom DNN hardware accelerator. The compiler uses model definition files created from popular frameworks to generate custom instructions. Different levels of hardware aware code optimizations are applied to improve performance and data reuse. The software also exposes an interface to run the accelerator implemented on various FPGA platforms, proving an end-to-end solution.

No description available.

In this article-based thesis, we cover applications of deep learning to different problems in condensed matter physics, where the goal is to either accelerate the computation or design of a nanoscale material. We first motivate and introduce how machine learning methods can be used to accelerate traditional condensed matter physics calculations. In addition, we discuss what designing a material means, and how it has been previously done. We then consider the fundamentals of electronic structure and conventional calculations which include density functional theory (DFT), density functional perturbation theory (DFPT), quantum Monte Carlo (QMC), and electron transport with tight binding. In addition, we cover the basics of deep learning. Afterwards, we discuss 6 articles. The first 5 articles are dedicated to accelerating the computation of nanoscale materials. In Article 1, we use convolutional neural networks to predict energies for diatomic molecules modelled with a Lennard-Jones potential and density functional theory energies of hexagonal lattices with and without defects. In Article 2, we use extensive deep neural networks to represent density functional theory energy functionals for electron gases by using the electron density as input and bypass the Kohn-Sham equations by using the external potential as input. In addition, we use deep convolutional inverse graphics networks to map the external potential directly to the electron density. In Article 3, we use voxel deep neural networks (VDNNs) to map electron densities to kinetic energy densities and functional derivatives of the kinetic energies for graphene lattices. We also use VDNNs to calculate an electron density from a direct minimization calculation and introduce a Monte Carlo based solver that avoids taking a functional derivative altogether. In Article 4, we use a deep learning framework to predict the polarization, dielectric function, Born-effective charges, longitudinal optical transverse optical splitting, Raman tensors, and Raman spectra for 2 crystalline systems. In Article 5, we use VDNNs to map DFT electron densities to QMC energy densities for graphene systems, and compute the energy barrier associated with forming a Stone-Wales defect. In Article 6, we design a graphene-based quantum transducer that has the ability to physically split valley currents by controlling the pn-doping of the lattice sites. The design is guided by an neural network that operates on a pristine lattice and outputs a lattice with pn-doping such that valley currents are optimally split. Lastly, we summarize the thesis and outline future work.

The aim of the thesis was to create a tool for time-series prediction based on deep learning. The first part of the work is a brief description of deep learning and its comparison to classical machine learning. In the next section contains brief analysis of some tools, that are already used for time-series forecasting. The last part is focused on the analysis of the problem as well as on the actual creation of the program.