Haze and smog are among the most common environmental factors impacting image quality and, therefore, image analysis. In this paper, I propose an end-to-end generative method for single image dehazing problem. It is based on fully convolutional network and effective network structures to recognize haze structure in input images and restore clear, haze-free ones. The proposed method is agnostic in the sense that it does not explore the atmosphere scattering model, it makes use of convolutional networks advantage in feature extraction and transfer instead. Somewhat surprisingly, it achieves superior performance relative to all existing state-of-the-art methods for image dehazing even on SOTS outdoor images, which are synthesized using the atmosphere scattering model. In order to improve its weakness in indoor hazy images and enhance the dehazed image's visual quality, a lightweight parallel network is put forward. It employs a different convolution strategy that extracts features with larger reception field to generate a complementary image. With the help of a parallel stream, the fusion of the two outputs performs better in PSNR and SSIM than other methods. / Thesis / Master of Applied Science (MASc)

We propose an end-to-end trainable Convolutional Neural Network (CNN), named GridDehazeNet, for single image dehazing. The GridDehazeNet consists of three modules: pre-processing, backbone, and post-processing. The trainable pre-processing module can generate learned inputs with better diversity and more pertinent features as compared to those derived inputs produced by hand-selected pre-processing methods. The backbone module implements a novel attention-based multi-scale estimation on a grid network, which can effectively alleviate the bottleneck issue often encountered in the conventional multi-scale approach. The post-processing module helps to reduce the artifacts in the final output. Experimental results indicate that the GridDehazeNet outperforms the state-of-the-art on both synthetic and real-world images. The proposed hazing method does not rely on the atmosphere scattering model and we provide an explanation as to why it is not necessarily beneficial to take advantage of the dimension reduction offered by the atmosphere scattering model for image dehazing, even if only the dehazing results on synthetic images are concerned. / Thesis / Master of Applied Science (MASc)

In the simple form, a communication system includes a transmitter and a receiver. In the transmitter, it transforms the one-hot vector message to produce a transmitted signal. In general, the transmitter demands restrictions on the transmitted signal. The channel is defined by the conditional probability distribution function. On receiving of the transmitted signal with noise, the receiver appears to apply the transformation to generate the estimate of one hot vector message. We can regard this simplest communication system as a specific case of autoencoder from a deep learning perspective. In our case, autoencoder used to learn the representations of the one-hot vector which are robust to the noise channel and can be recovered at the receiver with the smallest probability of error. Our task is to make some improvements on the autoencoder systems. We propose different schemes depending on the different cases. We propose a method based on optimization of softmax and introduce the L1/2 regularization in MSE loss function for SISO case and MIMO case, separately. The simulation shows that both our optimized softmax function method and L1/2 regularization loss function have a better performance than the original neural network framework. / Thesis / Master of Applied Science (MASc)

Memoria para optar al título de Ingeniero Civil Eléctrico / La enfermedad de Parkinson es un trastorno neurodegenerativo y objeto de interés en la comunidad científica, dado que existe una mayor incidencia luego de los 60 años en promedio y además la población mundial posee una alta esperanza de vida, implicando que existirá una mayor cantidad de casos en el futuro. Los tratamientos actuales para este trastorno se componen de un generador de pulsos yelectrodos, en donde se estimula de manera constante una zona del cuerpo (puede ser un área cerebral o parte de la médula) independiente del estado actual del paciente, lo cual conlleva en algunos efectos secundarios no deseados. El presente trabajo de memoria se enfoca en esta problemática, ya que busca un lazode control para estos tratamientos y corresponde a una de las primeras aproximaciones con aprendizaje de máquinas. En particular, se estudia un clasificador de movimiento con incertidumbre a partir de la actividad neural de un modelo animal de rata de 6-OHDA. De esta manera, se realiza la extracción de movimiento a partir de un video y de señales cerebrales de un modelo animal de Parkinson a través de algoritmos de ventanas deslizantes, generando imágenes de potencia en ciertas bandas de frecuencia con una etiqueta respectiva de movimiento a partir del video. Estas imágenes sirven para entrenar a un clasificador de Deep Learning Bayesiano, el cual puede extraer incertidumbre en la clasificación. Así, al utilizar Deep Learning Bayesiano con la forma de evaluación de MC Dropout se llega a obtener un recall de 80 % para la etiqueta de movimiento y la base de datos consistente en una ventana deslizante de medio segundo. Además, esta arquitectura es superior (para esta base de datos) en comparación a la de Deep Learning y evaluación estándar de dropout. Por otro lado, para estos resultados se tiene que con una mejor clasificación se obtiene una menor incertidumbre, lo cual es una de las ventajas al usar Deep Learning Bayesiano pues permite obtener una medida de confianza en la clasificación al realizar evaluaciones estocásticas. Finalmente, cabe destacar que este trabajo puede usarse como base para obtener una estrategia de control para un circuito cerrado específico para cada paciente, el cual posee incertidumbre en predicciones implicando en la confianza que posee el sistema para cambiar un estado específico. Para generar una nueva estrategia más robusta con incertidumbre, se debería repetir este experimento agregando nuevos biomarcadores o indicadores fisiológicos, además de explorar otros algoritmos para extracción de movimiento para el etiquetado de la base de datos. / FONDECYT

Indiana University-Purdue University Indianapolis (IUPUI) / Current research in deep learning is primarily focused on using Python as a support language. Go, an emerging language, that has many benefits including native support for concurrency has seen a rise in adoption over the past few years. However, this language is not widely used to develop learning models due to the lack of supporting libraries and frameworks for model development. In this thesis, the use of Go for the development of neural network models in general and convolution neural networks is explored. The proposed study is based on a Go-CUDA implementation of neural network models called GoCuNets. This implementation is then compared to a Go-CPU deep learning implementation that takes advantage of Go's built in concurrency called ConvNetGo. A comparison of these two implementations shows a significant performance gain when using GoCuNets compared to ConvNetGo.

<p>  </p> <p>The abundance of heterogeneous data produced and collected each day via multimodal sources may contain hidden events of interest, but in order to extract them the streams of data need to be analyzed with appropriate algorithms, so these events are presented to the end user at the right moment and at the right time. This dissertation proposes a series of algorithms that shape a comprehensive framework for situational knowledge on demand to address this problem. The framework consists of several modules and approaches, each of them is presented in a separate chapter: I begin with video data analysis in streaming video and video at rest for enhanced object detection of real-life surveillance video. For detecting the rare events of interest, I develop a semantic video analysis algorithm which uses an overlay knowledge graph and a semantical network. I show that the usage of the external knowledge for understanding the semantic analysis outperforms other techniques such as transfer learning. </p> <p>The semantical outliers can be used further for improving the algorithm of detecting new objects in the stream of different modalities. I extend the framework with additional modules for natural language data and apply the extended version of the semantic analysis algorithm to define the events of interest from multimodal streaming data. I present a way of combining several feature extractors which can be extended to multiple heterogeneous streams of data in order to efficiently fuse the data based on its semantical similarity, and then show how the serverless architecture of the framework outperforms conventional cloud software architecture. </p> <p>Besides detecting the rare and semantically incompatible events, the semantic analysis can be used for improving the neural networks performance with the data augmentation. The algorithm presented for augmenting the data with the potentially novel objects to circumvent the data drift problem uses the knowledge graph and generative adversarial networks to present the objects to augment the training datasets for supervised learning. I extend the presented framework with a pipeline for generating synthetic novelties to improve the performance of feature extractors and provide the empirical evaluation of the developed method.</p>

The camera is the cheapest and computationally real-time option for detecting or segmenting the environment for an autonomous vehicle, but it does not provide the depth information and is undoubtedly not reliable during the night, bad weather, and tunnel flash outs. The risk of an accident gets higher for autonomous cars when driven by a camera in such situations. The industry has been relying on LiDAR for the past decade to solve this problem and focus on depth information of the environment, but LiDAR also has its shortcoming. The industry methods commonly use projections methods to create a projection image and run detection and localization network for inference, but LiDAR sees obscurants in bad weather and is sensitive enough to detect snow, making it difficult for robustness in projection based methods. We propose a novel pointwise and Instance segmentation deep learning architecture for the point clouds focused on self-driving application. The model is only dependent on LiDAR data making it light invariant and overcoming the shortcoming of the camera in the perception stack. The pipeline takes advantage of both global and local/edge features of points in points clouds to generate high-level feature. We also propose Pointer-Capsnet which is an extension of CapsNet for small 3D point clouds.

No description available.

Robust Speaker Recognition (SR) has been a focus of attention for researchers since long. The advancement in speech-aided technologies especially biometrics highlights the necessity of foolproof SR systems. However, the performance of a SR system critically depends on the quality of speech features used to represent the speaker-specific information. This research aims at extracting the speaker-specific information from Mel-frequency Cepstral Coefficients (MFCCs) using deep learning. Speech is a mixture of various information components that include linguistic, speaker-specific and speaker’s emotional state information. Feature extraction for each information component is inevitable in different speech-related tasks for robust performance. However, almost all forms of speech representation carry all the information as a whole, which is responsible for the compromised performances by SR systems. Motivated by the complex problem solving ability of deep architectures by learning high-level task-specific information in the data, we propose a novel Deep Neural Architecture (DNA) to extract speaker-specific information (SI) from MFCCs, a popular frequency domain speech signal representation. A two-stage learning strategy is adopted, which is based on unsupervised training for network initialisation followed by regularised contrastive learning. To train our network in the 2nd stage, we devise a contrastive loss function to discriminate the speakers on the basis of their intrinsic statistical patterns, distributed in the representations yielded by our deep network. This is achieved in the contrastive pair-wise comparison of these representations for similar or dissimilar speakers. To improve the generalisation and reduce the interference of environmental effects with the speaker-specific representation, we regulate the contrastive loss with the data reconstruction loss in a multi-objective optimisation. A detailed study has been done to analyse the parametric space in training the proposed deep architecture for optimum performance. Finally we compare the performance of our learned speaker-specific representations with several state-of-the-art techniques in speaker verification and speaker segmentation tasks. It is evident that the representations acquired through learned DNA are invariant and comparatively less sensitive to the text, language and environmental variability.

Capsule Networks are the next generation of image classifiers. Although they have several advantages over conventional Convolutional Neural Networks (CNNs), they remain computationally heavy. Since inference on Capsule Networks is timeconsuming, thier usage becomes limited to tasks in which latency is not essential. Approximation methods in Deep Learning help networks lose redundant parameters to increase speed and lower energy consumption. In the first part of this work, we go through an algorithm called zero-skipping. More than 50% of trained CNNs consist of zeros or values small enough to be considered zero. Since multiplication by zero is a trivial operation, the zero-skipping algorithm can play a massive role in speed increase throughout the network. We investigate the eligibility of Capsule Networks for this algorithm on two different datasets. Our results suggest that Capsule Networks contain enough zeros in their Primary Capsules to benefit from this algorithm. In the second part of this thesis, we investigate pruning as one of the most popular Neural Network approximation methods. Pruning is the act of finding and removing neurons which have low or no impact on the output. We run experiments on four different datasets. Pruning Capsule Networks results in the loss of redundant Primary Capsules. The results show a significant increase in speed with a minimal drop in accuracy. We also, discuss how dataset complexity affects the pruning strategy. / Graduate