Learning discriminative feature representations has attracted a great deal of attention since it is a critical step to facilitate the subsequent classification, retrieval and recommendation tasks. In this dissertation, besides incorporating prior knowledge about image labels into the image classification as most prevalent feature learning methods currently do, we also explore some other general-purpose priors and verify their effectiveness in the discriminant feature learning. As a more powerful representation can be learned by implementing such general priors, our approaches achieve state-of-the-art results on challenging benchmarks. We elaborate on these general-purpose priors and highlight where we have made novel contributions. We apply sparsity and hierarchical priors to the explanatory factors that describe the data, in order to better discover the data structure. More specifically, in the first approach we propose that we only incorporate sparse priors into the feature learning. To this end, we present a support discrimination dictionary learning method, which finds a dictionary under which the feature representation of images from the same class have a common sparse structure while the size of the overlapped signal support of different classes is minimised. Then we incorporate sparse priors and hierarchical priors into a unified framework, that is capable of controlling the sparsity of the neuron activation in deep neural networks. Our proposed approach automatically selects the most useful low-level features and effectively combines them into more powerful and discriminative features for our specific image classification problem. We also explore priors on the relationships between multiple factors. When multiple independent factors exist in the image generation process and only some of them are of interest to us, we propose a novel multi-task adversarial network to learn a disentangled feature which is optimized with respect to the factor of interest to us, while being distraction factors agnostic. When common factors exist in multiple tasks, leveraging common factors cannot only make the learned feature representation more robust, but also enable the model to generalise from very few labelled samples. More specifically, we address the domain adaptation problem and propose the re-weighted adversarial adaptation network to reduce the feature distribution divergence and adapt the classifier from source to target domains.

This thesis shows how children’s learning is influenced and modified by the teaching environment. The metacognitive, self-regulatory learning behaviours of sixteen kindergarten students were examined in order to determine how students perceive learning, either by adopting deep approaches, where the focus is on understanding and meaning, or surface approaches, where the meeting of institutional demands frequently subjugate the former goals. The data have been analysed within a qualitative paradigm from a phenomenographic perspective. The study addresses three issues: the nature and frequency of the strategic learning behaviours displayed by the students; the contribution strategic behaviours make to the adoption of deep or surface learning approaches; and how metacognitive teaching environments influence higher-order thinking. Findings reveal that where teachers had metcognitive training, the frequency of strategy use increased irrespective of student performance. High achieving students used more strategic behaviours, used them with greater efficiency, and tended to display more of the characteristics of deep approach learners. This study suggests that many of the differential outcomes evident amongst students may be substantially reduced through early and consistent training within a teaching environment conductive to the development of metacognitive, self-regulatory behaviours and deep learning approaches / Doctor of Philosophy (PhD)

Deep generative neural networks such as the Deep Belief Network and Deep Boltzmann Machines have been used successfully to model high dimensional visual data. However, they are not robust to common variations such as occlusion and random noise. In this thesis, we explore two strategies for improving the robustness of DBNs. First, we show that a DBN with sparse connections in the first layer is more robust to variations that are not in the training set. Second, we develop a probabilistic denoising algorithm to determine a subset of the hidden layer nodes to unclamp. We show that this can be applied to any feedforward network classifier with localized first layer connections. By utilizing the already available generative model for denoising prior to recognition, we show significantly better performance over the standard DBN implementations for various sources of noise on the standard and Variations MNIST databases.

Deep generative neural networks such as the Deep Belief Network and Deep Boltzmann Machines have been used successfully to model high dimensional visual data. However, they are not robust to common variations such as occlusion and random noise. In this thesis, we explore two strategies for improving the robustness of DBNs. First, we show that a DBN with sparse connections in the first layer is more robust to variations that are not in the training set. Second, we develop a probabilistic denoising algorithm to determine a subset of the hidden layer nodes to unclamp. We show that this can be applied to any feedforward network classifier with localized first layer connections. By utilizing the already available generative model for denoising prior to recognition, we show significantly better performance over the standard DBN implementations for various sources of noise on the standard and Variations MNIST databases.

This qualitative research study seeks to answer the question: Which instructional design approaches for blended learning encourage deep learning? This grounded theory research captures the lived experiences of instructional designers and faculty members in converting courses at the post-secondary level from traditional, face-to-face delivery to blended delivery using educational technology. Study results provide insight into the complexities involved in the design and development of blended delivery courses and shed light on the complications that can arise with course conversion. The study also opens a window into design approaches to foster deep learning, clarifying the importance of targeting high levels of learning in the course syllabus / outline, and then aligning every part of the course to the specific learning outcomes identified. Study results culminate in a set of recommended instructional design approaches that foster deep learning in a blended learning environment. / 2013-08

Face recognition, although being a popular area of research for over a decade has still many open research challenges. Some of these challenges include the recognition of poorly illuminated faces, recognition under pose variations and also the challenge of capturing sufficient training data to enable recognition under pose/viewpoint changes. With the appearance of cheap and effective multimodal image capture hardware, such as the Microsoft Kinect device, new possibilities of research have been uncovered. One opportunity is to explore the potential use of the depth maps generated by the Kinect as an additional data source to recognize human faces under low levels of scene illumination, and to generate new images through creating a 3D model using the depth maps and visible-spectrum/RGB images that can then be used to enhance face recognition accuracy by improving the training phase of a classification task. With the goal of enhancing face recognition, this research first investigated how depth maps, since not affected by illumination, can improve face recognition, if algorithms traditionally used in face recognition were used. To this effect a number of popular benchmark face recognition algorithms are tested. It is proved that algorithms based on LBP and Eigenfaces are able to provide high level of accuracy in face recognition due to the significantly high resolution of the depth map images generated by the latest version of the Kinect device. To complement this work a novel algorithm named the Dense Feature Detector is presented and is proven to be effective in face recognition using depth map images, in particular under wellilluminated conditions. Another technique that was presented for the goal of enhancing face recognition is to be able to reconstruct face images in different angles, through the use of the data of one frontal RGB image and the corresponding depth map captured by the Kinect, using faster and effective 3D object reconstruction technique. Using the Overfeat network based on Convolutional Neural Networks for feature extraction and a SVM for classification it is shown that a technically unlimited number of multiple views can be created from the proposed 3D model that consists features of the face if captured real at similar angles. Thus these images can be used as real training images, thus removing the need to capture many examples of a facial image from different viewpoints for the training of the image classifier. Thus the proposed 3D model will save significant amount of time and effort in capturing sufficient training data that is essential in recognition of the human face under variations of pose/viewpoint. The thesis argues that the same approach can also be used as a novel approach to face recognition, which promises significantly high levels of face recognition accuracy base on depth images. Finally following the recent trends in replacing traditional face recognition algorithms with the effective use of deep learning networks, the thesis investigates the use of four popular networks, VGG-16, VGG-19, VGG-S and GoogLeNet in depth maps based face recognition and proposes the effective use of Transfer Learning to enhance the performance of such Deep Learning networks.

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

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.