Deep Neural Networks (DNN)

I the undersigned do hereby endorse that the information contained herein is my original work under the management of my supervisor. Where necessary, citations have been added to acknowledge other researchers whose work I have referred to. Name Reg No.

This research study about image classification by using the deep neural network (DNN) or also known as Deep Learning by using framework TensorFlow. Python is used as a programming language because it comes together with TensorFlow framework. The input data mainly focuses in flowers category which there are five (5) types of flowers that have been used in this paper. Deep neural network (DNN) has been choosing as the best option for the training process because it produced a high percentage of accuracy. Results are discussed in terms of the accuracy of the image classification in percentage. For the test images, all flower images are correctly classified thus leading to the test accuracy of 100% for the proposed DNN model.

—This work proposes a low power digital block for spatial localization of sensors within the limited area/power budget of sensor nodes. We show a novel digital architecture for a Linear Program (LP) solver based on a recurrent (non-linear feedback) Neural Network (RNN). Training data is not needed in our approach. We solve the primal and dual optimization problems for spatial localization with a single multi-functional data path which does not require matrix inversions. FPGA and ASIC implementations are presented which target a sensor (microrobot) localization problem in 2D using angle-of-arrival (AOA) measurements. The results show that the estimated locations (2D coordinates) are very close to the ground truth values in all tested scenarios. The proposed RNN has input and output layers, and a hidden layer with four neurons. FPGA implementation of the localizer in 180 nm process dissipates 180 mW of power at 1.5V and 31.25 MHz. When scaled to 128 neurons the performance is 13 Mop/sec/W for the same FPGA technology. The design has been also simulated in an ASIC 45 nm (PDK45 1V VDD) standard cell technology. With 128 neurons in the hidden layer the ASIC consumes 196.9 mW power at 516 MHz, which is equivalent to 677.165 Mop/sec/W. The proposed design significantly improves computing efficiency over the recently published results for similar implementations.

Improving phoneme recognition has attracted the attention of many researchers due to its applications in various fields of speech processing. Recent research achievements show that using deep neural network (DNN) in speech recognition systems significantly improves the performance of these systems. There are two phases in DNN-based phoneme recognition systems including training and testing. Most previous research attempted to improve training phase such as training algorithms, different types of network, network architecture, feature type, etc. But in this study, we focus on test phase which is related to generate phoneme sequence that is also essential to achieve good phoneme recognition accuracy. Past research used Viterbi algorithm on hidden Markov model (HMM) to generate phoneme sequences. We address an important problem associated with this method. To deal with the problem of considering geometric distribution of state duration in HMM, we use real duration probability distribut...

The question of noise may not seem immediately relevant to recent upheavals in the political economy of music. But since I was asked to speak about streaming at the 'Ebbing Sounds' conference that zweikommasieben co-hosted last May in San Francisco and since my research is focused on the concept of noise, I put together some thoughts on that basis. Noise has a specific sharply defined computational meaning in several different theoretical frameworks in science, most centrally information theory, where it is defined in terms of uncertainty and interference. It also has a variety of social and cultural understandings that are more open to interpretation, related to perceptual and affective states and to general notions like randomness, unpredictability, disorder, resistance, and excess. Many of these concepts – information, randomness, noise – which have very specific scientific and mathematical meanings, are routinely employed to describe phenomena at the cultural level. There, they are at best problematically overextended, and more often serve as tools in ideologically motivated justifications for the continuing march of neoliberalism's creative destruction. The central claim of my work on noise is pretty straightforward: noise is relative to theoretical framework, level of description, and scale of analysis. Its well known that first the printing press, then early sound recording and transmission technologies – phonograph and radio – induced a scale transformation in the political economy of music, and that digital sound, file sharing, and now streaming have given rise to similarly radical transitions. Each of these technical developments profoundly altered the sonic domain, changing not only the distribution of music but also its composition and production, the ecology of listening, and the understanding of noise. This is also the case in the recent consolidation of the streaming economy, which needs to be thought of in terms of platform capitalism and the transition from ICT to AI. This is a shift from centralized broadcast media to interactive media, user generated content, and targeted advertising. But it is also a concurrent shift in the management of noise, as relatively fixed exclusionary frameworks based on robustness to noise give way to data-driven fluid modulation of relevance and resilience to noise. Lossy compression formats like the mp3, which made file-sharing and now streaming possible, are based on perceptual coding – which uses a model of the auditory system to compress a signal by exploiting redundancies in the way that auditory information is processed. Like noise, compression is relative to 1 theoretical framework: that is, the adequacy of a compressed representation to its object, or its utility for a subject, is dependent on the goal that it serves and the context in which it arises. In general it is computational resource allocation problems that dictate the pertinent level of compression in a given economy, whether cognitive-perceptual, or digital-financial. It is common knowledge that the huge impact of the mp3 format was due to achieving a reduction in the computational cost of storing and transmitting auditory data. But we can also talk of evolutionary biology as a process of generating compact representations, or of perception as a form of lossy source compression. This generalized notion of 2 compression is precisely the model of thought underlying recent theoretical developments in AI and cognitive science, a model that is undoubtedly now driving decisions in boardroom meetings at tech startups. Understanding political economy in general and platform capitalism in particular requires engagement with the problem of naturalization (explanation according to natural laws). Mainstream, or orthodox, economic theory is a naturalization of economic activity according to principles taken from physics. This is a brutal abstraction relying on a number of simplifying assumptions, so it treats various details as noise (most notably mainstream economics forgets asymmetries, historicity, and communication). More recent heterodox theories move to a biological or ecological paradigm, which accounts for some non-equilibrium 1 Sterne, J. (2012) MP3 – The Meaning of a Format. Duke University Press.

Deep Learning is becoming a very popular subset of machine learning due to its high level of performance across many types of data. A great way to use deep learning to classify images is to build a convolutional neural network (CNN). The Tensorflow library in Python makes it pretty simple to build a CNN. Our main purpose of this experiment, training biscuits and seperate them two categories. Shape and Type of made.In our experiment we have two shapes of biscuits, round and rectangular.We also have two types such as Cocoa Biscuits and Normal (Milky) Biscuits. In our experiment we start with collecting data and labeling after that we classify images and train that with position (label) argument. So our last stage is Real-time classify images with open-cv library.

The field of Computer Vision is a branch of science of the computers and systems of software in which one can visualize and as well as comprehend the images and scenes given in the input. This field is consisting of numerous aspects for example image recognition, the detection of objects, generation of images, image super resolution and more others. Object detection is broadly utilized for the detection of faces, the detection of vehicles, counting of pedestrians on a certain street, images displayed on the web, security systems and cars with the feature of self-driving. This process also encompasses the precision of every technique for recognizing the objects. The detection of objects is a crucial task; however, it is also a very challenging vision task. It is an analytical subdivide of various applications such as searching of images, image auto-annotation or scene understanding and tracking of various objects. The tracking of objects in motion of video image sequence was one of the most important subjects in computer vision.

Computer vision learning pays off attention due to the global epidemic COVID-19 to improve public safety. Since 2019
there has been a huge human loss cause of COVID-19 and the world is at risk. When it comes to human safety technology has
always played a pivotal role to protect earth and its humanity from various calamities. In this era of COVID-19, Wearing masks
is a very important thing for further transmission of this deadly virus from one person to another. But many of us are avoiding
wearing masks in public places, to overcome this issue we have developed face mask detection software that detects if a person is
wearing mask or not. Installing this software module in public places, in work environment and at various crowded areas will
ease the surveillance of face detection than manual surveillance by man power.
Deep neural networks have always shown highly successful results when it comesto object detection.
MOBILENETV2 is one such perfect model to be used in facemask detection. Moreover, face mask detection done by
MobileNetV2 is very much better when measuring real-time execution in relation to powerful GPU. And the ability to calculate
with low memory MobileNetV2 command it is now possible for us to detect faces in real-time mode. The processing strength with
low power memory MobileNetV2 command is acceptable for real-time processing. In our research project we have focused on
detecting people with face mask and without facemask in real-time. Under the technical and practical conditions, the complete
real-time video data detection is finalized with over-availability, localization, and recognition methods. Test results show normal
loss is 0.0730 after training 20 epochs. After training 20 points for epochs MAP is 0.96. This is new unique Method of face
maskingsystem visual output with 96% categories and accurate detection.

This study and research aimed is to classify and predict the credit card default customers payment by means of contemporary approach of artificial neural network (ANN) known as deep neural network. This paper explains the dataset which signifies Taiwan credit card defaults in 2005 and their previous payment histories taken from popular machine learning dataset resource known as UCI. The paper enlightens each and every concept and step require to build, train, validate and test a deep neural network model for classification task that has never been discussed before. Moreover, we tried to elaborate the relevant and important concepts associated with deep neural network model that must be kept in mind during model building. This paper mainly tries to classify the default payment customer with more than 82% accuracy. For this purpose, various deep neural network techniques with different libraries are used to attain maximum accuracy and we have tried to build a best possible model which can be used for future prediction. This study proves deep neural network is the only one that can accurately estimate the real probability of default. So, by using this network model, which is more complex, sophisticated and most widely used than a simple neural network and logistic regression model, the classification simulation shall have a better performance and accuracy.

En el periodo de la democratizacion de la informacion y la masicacion de datos, donde
cada aparato electronico genera datos de varios tipos y estructuras, sean estos sobre nuestra posicion geografica, nuestros gustos, nuestra informacion personal, familiar, profesional, entre otros. La curiosidad humana ha llegado mas lejos de lo que hace unas decadas atras se hubiese imaginado, de manera similar el aumento y mejora constante de la potencia del hardware de procesamiento para esta masificacion de datos, nos ha llevado a utilizar algoritmos de manera mas eficiente a cuando estos algoritmos fueron desarrollados, llevandonos a procesar millones datos historicos para predecir el comportamiento futuro de algun sistema natural o articial, esto ha motivado la realizacion de este trabajo, el cual consiste en definitiva evitar colisiones o accidentes en los vehiculos, prediciendo los puntos geograficos donde ocurriran estas colisiones, con la ayuda de datos previos de ubicacion GPS, de los accidentes y las redes neuronales profundas DNN por sus siglas en ingles. El desarrollo de este trabajo tiene como base el estudio realizado en el paper \Predicting Traffic Accidents Through Heterogeneous UrbanData: A Case Study". (Zhuoning Yuan, 2017), muestra exploraciones de tecnicas efectivas para abordar los desafios de la prediccion de accidentes de trafico, y obtener mejores resultados de prediccion, lo formula como un problema de clasicacion binaria. Para cada segmento de la carretera en cada hora, predice si ocurrira o no un accidente, para ello compara varios algoritmos, demostrando que el modelo que mejor funciona es el DNN. Los autores del paper consideran muchas caracteristicas en su modelo, incluyen varios atributos de tiempo, clima, carretera y factores humanos, para este trabajo se consideran pocas caracterirsticas, se utiliza el DNN modelado en el paper original, ademas se utiliza el metodo geohashes para agrupar la geo ubicacion de los incidentes. Se utiliza la arquitectura de red completamente conectada con 3 capas ocultas, con las conguraciones predeterminadas sugeridas por los autores, utilizando dropout regularization para prevenir sobreajustes, y una salida binaria.

It is a difficult task of continuous automatic speech recognition, translating of spoken words into text due to the excessive viability in speech signals. In recent years speech recognition has been accomplishing pinnacle of success however it still has few limitations to overcome. Deep learning also known as representation learning or sometimes referred as unsupervised feature learning, is a subset of machine learning. Deep learning is becoming a conventional technology for speech recognition and has efficiently replaced Gaussian mixtures for speech recognition on a global scale. The predominant goal of this undertaking is to apply deep learning algorithms, together with Deep Neural Networks (DNN) and Deep Belief Networks (DBN), for automatic non-stop speech recognition. Keywords: Gaussian Mixture Model (GMM), Hidden Markov Models (HMMs), Deep Neural Networks (DNN), Deep Belief Networks (DBN).

Network pruning is a widely used technique to reduce computation cost and model size for deep neural networks. However, the typical three-stage pipeline, i.e., training, pruning and retraining (finetuning) significantly increases the overall training trails. For instance, the retraining process could take up to 80 epochs for ResNet-18 on ImageNet, that is 70% of the original model training trails. In this paper, we develop a systematic weight-pruning optimization approach based on Surrogate Lagrangian relaxation (SLR), which is tailored to overcome difficulties caused by the discrete nature of the weight-pruning problem while ensuring fast convergence. We decompose the weight-pruning problem into subproblems, which are coordinated by updating Lagrangian multipliers. Convergence is then accelerated by using quadratic penalty terms. We evaluate the proposed method on image classification tasks, i.e., ResNet-18 and ResNet-50 using ImageNet, and ResNet-18, ResNet-50 and VGG-16 using CIF...

Diagnosis of heart disease is a difficult job, and researchers have designed various intelligent diagnostic systems for improved heart disease diagnosis. However, low heart disease prediction accuracy is still a problem in these systems. For better heart risk prediction accuracy, we propose a feature selection method that uses a floating window with adaptive size for feature elimination (FWAFE). After the feature elimination, two kinds of classification frameworks are utilized, i.e., artificial neural network (ANN) and deep neural network (DNN). Thus, two types of hybrid diagnostic systems are proposed in this paper, i.e., FWAFE-ANN and FWAFE-DNN. Experiments are performed to assess the effectiveness of the proposed methods on a dataset collected from Cleveland online heart disease database. The strength of the proposed methods is appraised against accuracy, sensitivity, specificity, Matthews correlation coefficient (MCC), and receiver operating characteristics (ROC) curve. Experime...

MOOC represents an ultimate way to deliver educational content in higher education settings by providing high-quality educational material to the students throughout the world. Considering the differences between traditional learning paradigm and MOOCs, a new research agenda focusing on predicting and explaining dropout of students and low completion rates in MOOCs has emerged. However, due to different problem specifications and evaluation metrics, performing a comparative analysis of state-of-the-art machine learning architectures is a challenging task. In this paper, we provide an overview of the MOOC student dropout prediction phenomenon where machine learning techniques have been utilized. Furthermore, we highlight some solutions being used to tackle with dropout problem, provide an analysis about the challenges of prediction models, and propose some valuable insights and recommendations that might lead to developing useful and effective machine learning solutions to solve the MOOC dropout problem.

Hardware-based machine learning is becoming increasingly popular due to its high speed of computation. One of the desired characteristics of such hardware is reduced hardware and design costs. This paper proposes a design approach for a neural network to reduce the cost of hardware in terms of adders and multipliers. Adders and multipliers are parts of the main components in the neural network, and they are used in each node in the network. The proposed approach reduces the number of multipliers and adders in the network by half, which reduces the cost. The proposed technique is based on sharing multiplier and adder between two hidden layers. The method has been tested and validated using multiple datasets. The accuracy of the proposed approach is similar to the traditional methods in the literature, while the proposed approach utilizes only half the number of multipliers and adders. The proposed design is implemented using VHDL and Altera Arria 10 GX FPGA. The simulation result shows the proposed method retains the performance of the network with a 63% reduction in the hardware design with acceptable accuracy.

Deep learning and image processing are two areas of great interest to academics and industry professionals alike. The areas of application of these two disciplines range widely, encompassing fields such as medicine, robotics, and security and surveillance. The aim of this book, 'Deep Learning for Image Processing Applications', is to offer concepts from these two areas in the same platform, and the book brings together the shared ideas of professionals from academia and research about problems and solutions relating to the multifaceted aspects of the two disciplines. The first chapter provides an introduction to deep learning, and serves as the basis for much of what follows in the subsequent chapters, which cover subjects including: the application of deep neural networks for image classification; hand gesture recognition in robotics; deep learning techniques for image retrieval; disease detection using deep learning techniques; and the comparative analysis of deep data and big data. The book will be of interest to all those whose work involves the use of deep learning and image processing techniques

In telecommunications, 6G is the sixth generation standard currently under development for wireless communications technologies supporting cellular data networks. It is the planned successor to 5G and will likely be significantly faster. Like its predecessors, 6G networks will probably be broadband cellular networks, in which the service area is divided into small geographical areas called cells. Several companies (Nokia, Ericsson, Huawei, Samsung, LG, Apple), as well as several countries (China, South Korea, and Japan), have shown interest in 6G networks.

The system of using pre-made bar codes to identify a product during its billing process is time-consuming and labour intensive. The relatively unique barcode needs to be first produced, then it must be manually attached to the product. This requires a lot of pre-processing work on the products to make them ready for identification and classification. This paper presents an alternate system that works on the principle of using the products' natural characteristics like its discrete and distinguishable looks to identify and classify them during the billing process. It's mimicking the human way of identifying and distinguishing the products. To implement this system, we need to move away from conventional methods of programming and use a different paradigm for designing software systems based on an artificial intelligence concept i.e., machine learning. We use machine learning techniques to design the working philosophy of this system. The algorithms in Deep Neural Networks which is a part of Artificial Neural Networks, help in creating a model to base our software system's operational logic. Especially the models based on Convolutional Neural Networks have been proven to be efficient in providing a model for image classification. This paper discusses the abstracted software system from the base billing process without worrying about the hardware environment. We choose Python and its web framework Django to design the UI to implement this system over a distributed network within any establishment that needs to incorporate this process so that each node that has to process billing need not have to adhere to the hardware requirements imposed on them to run the various CNN models which are reliant on the GPU-based tensor architecture of TensorFlow. The system also provides mechanisms for inventory management over distributed networks and simple data analytics based on local sales.

Scope of the book:
This book focusses on the technical concepts of deep learning and its associated branch Neural Networks for the various dimensions of image processing applications. The proposed volume intends to bring together researchers to report the latest results or progress in the development of the above-mentioned areas. Since there is a deficit of books on this specific subject matter, the editors aim to provide a common platform for researchers working in this area to exhibit their novel findings.
Topics of Interest:
This book solicits contributions, which include the fundamentals in the field of Deep Artificial Neural Networks and Image Processing supported by case studies and practical examples. Each chapter is expected to be self-contained and to cover an in-depth analysis of real life applications of neural networks to image analysis.

Lorsque l'on souhaite améliorer le confort thermique d'un bâtiment ancien, il est nécessaire de bien tenir compte de son fonctionnement hygrothermique spécifique, au risque de lui faire perdre ses qualités bioclimatiques intrinsèques et altérer sa structure.

MOOC represents an ultimate way to deliver educational content in higher education settings by providing high-quality educational material to the students throughout the world. Considering the differences between traditional learning paradigm and MOOCs, a new research agenda focusing on predicting and explaining dropout of students and low completion rates in MOOCs has emerged. However, due to different problem specifications and evaluation metrics, performing a comparative analysis of state-of-the-art machine learning architectures is a challenging task. In this paper, we provide an overview of the MOOC student dropout prediction phenomenon where machine learning techniques have been utilized. Furthermore, we highlight some solutions being used to tackle with dropout problem, provide an analysis about the challenges of prediction models, and propose some valuable insights and recommendations that might lead to developing useful and effective machine learning solutions to solve the ...

Deep learning is progressively gaining popularity as a viable alternative to i-vectors for speaker recognition. Promising results have been recently obtained with Convolutional Neural Networks (CNNs) when fed by raw speech samples directly. Rather than employing standard hand-crafted features, the latter CNNs learn low-level speech representations from wave-forms, potentially allowing the network to better capture important narrow-band speaker characteristics such as pitch and formants. Proper design of the neural network is crucial to achieve this goal. This paper proposes a novel CNN architecture, called SincNet, that encourages the first convolutional layer to discover more meaningful filters. SincNet is based on parametrized sinc functions, which implement band-pass filters. In contrast to standard CNNs, that learn all elements of each filter, only low and high cutoff frequencies are directly learned from data with the proposed method. This offers a very compact and efficient way to derive a customized filter bank specifically tuned for the desired application. Our experiments, conducted on both speaker identification and speaker verification tasks, show that the proposed architecture converges faster and performs better than a standard CNN on raw waveforms.

Wireless networks have evolved over the years and they have become some of the most prominent communication media. These networks generally transmit large volumes of information at any given time. This has engendered a number of security threats and privacy concerns. This paper presents a Deep Long Short-Term Memory (DLSTM) based classifier for wireless intrusion detection system (IDS). Using the NSL-KDD dataset, we compare the DLSTM IDS to existing methods such as Deep Feed Forward Neural Networks, Support Vector Machines, k-Nearest Neighbors, Random Forests and Naïve Bayes. The experimental results suggest that the DLSTM IDS outperformed existing approaches.

Engineering science is widely accustomed to detect the movement of lips. The data generated through visual motion of
mouth and corresponding audio are highly correlated. This fact has been exploited for lip reading and for improving speech
recognition. A CNN(Convolutional Neural Network) shall detect the movement of lips and judge the words spoken. This trained
CNN detects the words that are spoken within the video and displayed within the text format. The CNN also relies on
information provided by the context, knowledge of the language, and any residual hearing. The aim is to verify whether the
utilization of engineering science methods, namely DNN(Deep Neural Network), could also be an appropriate candidate for
solving this problem. within the sensible part, the most target is on presenting the results both in terms of the accuracy of the
trained neural network on test data.

Speech enhancement is the process of estimating the clean speech from the noisy speech signal in order to improve its quality or intelligibility. This is still a challenging problem in the field of speech processing. Some of the application areas for speech enhancement are communication, hearing aids, automatic speech recognition, speech coding, forensics and restoration of historic recordings.
This thesis proposes, implements and analyzes various speech sound class-specific and noise-specific enhancement approaches and frame-wise selection methods for class-specific and noise-specific models. 
1. We have analyzed the performance of our enhancement scheme, where we use various speech-sound class-specific dictionaries to enhance noisy speech. By removing the contribution from the bases of those classes that correlate well with noise, one could improve the enhancement performance. We achieve this by learning different dictionaries for different classes and select a particular dictionary for a frame. We explore a class-specific enhancement approach, where we use a sparse coding, dictionary-based approach to learn dictionaries of various speech classes and noises. When we analyze the performance of our various class-specific approaches in terms of phoneme recognition, we obtain performances superior to class-independent case, even when we use estimated (approximate) labels for enhancement.
2. An error in the estimated class labels in the class-specific approach results in the selection of erroneous dictionaries for enhancement. The joint enhancement-decoding (JED) algorithm that we propose tries to overcome this issue by jointly optimizing for labels of all the frames and the decoding path to improve the phoneme recognition accuracy. The current noisy speech frame is enhanced by multiple (N) phoneme-specific dictionaries close to the approximate label of that frame. These N enhanced frames are then fed into the JED algorithm. The algorithm accepts these N observations and chooses the best for each frame such that the overall likelihood is maximized to obtain the final recognized labels. The Viterbi decoding algorithm used in speech recognition is integrated with the class label selection to develop the JED algorithm.
3. We also propose a method of picking the best DNN model in the scenario, where multiple noise-specific DNN models are available for enhancement, using the Monte Carlo (MC) dropout proposed by Gal and Ghahramani. The variance measure of the output signal vectors, resulting from different MC dropout trials, is used as a measure of the model precision to select one out of the multiple models for each frame. We find this method to be particularly useful for unseen noisy scenario, where the noise corrupting the test speech is different from those with which the available DNN models are trained. For the unseen noisy scenario, this method performs better than selecting the model using a DNN classifier. We observe some promising results with the enhancement performance of the algorithm on speech corrupted with a mixture of multiple noises and for the case, where different segments of speech are corrupted by different noises.

In this paper, we present the design, implementation and evaluation of non-line-of-sight (NLOS) perception to achieve a virtual see-through functionality for road vehicles. In this system, a safety event, such as pedestrian crossing or traffic list status or vehicle merge, that are occluded to a driver due to another vehicle on the driving lane, are perceived and communicated by the occluding vehicle. Each vehicle is equipped with a camera, placed on the dashboard, that perceives the scene in a drivers view. This scene is analyzed and mapped into specific warning codes that are specific to a safety event, and communicated as short packets using visible light communication. The camera in the following vehicle captures this information and generates a recommendation for safety action to the driver by comparing the warning from the packet and from its own scene perception. Through experimental evaluations of a proof-of-concept implementation, we show that our system is able to achieve up to 90% accuracy in identifying nine occluded safety events, which correspond to traffic light statuses (red, green, yellow), other vehicles' lane change behaviors (merge/leave lane left/right), and pedestrian detection.

Speaker Recognition has been one of the most interesting yet challenging problem in the field of machine learning and artificial intelligence. It is used In areas of human voice authentication for security purpose and identifying a person from a group of speakers. It has been a grinding task to teach a machine the differences in human voices when people belong to different background like gender, language, accent.etc. In this paper, we use the deep learning approach to build and train two models : ANN and CNN; and compare their results. In former, the neural networks are fed on diverse extracted features from audio collection. The latter is a convolutional neural network trained on spectrograms. At last, we use transfer learning approach on both to get a viable output using less data.

MOOC represents an ultimate way to deliver educational content in higher education settings by providing high-quality educational material to the students throughout the world. Considering the differences between traditional learning paradigm and MOOCs, a new research agenda focusing on predicting and explaining dropout of students and low completion rates in MOOCs has emerged. However, due to different problem specifications and evaluation metrics, performing a comparative analysis of state-of-the-art machine learning architectures is a challenging task. In this paper, we provide an overview of the MOOC student dropout prediction phenomenon where machine learning techniques have been utilized. Furthermore, we highlight some solutions being used to tackle with dropout problem, provide an analysis about the challenges of prediction models, and propose some valuable insights and recommendations that might lead to developing useful and effective machine learning solutions to solve the ...

Transfer learning (TL), have been becoming extremely popular due to its appalling property of being able to work at a different domain, where it's not trained. Often training at such domain is very costly due to various reasons such as lack of sufficient data etc. Recently Deep Neural Network (DNN) based models have been widely used in the context of TL. These networks as clear transferable layers representing common latent features, hence intuitive to transfer. In this paper we survey various DNN architectures, different training methods adopted to train those and their performance in several application domains.

This paper investigates the effect of speaking rate variation on the task of frame classification. This task is indicative of the performance on phoneme and word recognition and is a first step towards designing voice-controlled interfaces. Different speaking rates cause different dynamics. For example, speaking rate variations will cause changes both in formant frequencies and in their transition tracks. A word spoken at normal speed gets recognized more often than the same word spoken by the same speaker at a much faster or slower pace, or vice-versa. It is thus imperative to design interfaces which take into account different speaking variabilities. To better incorporate speaker variability into digital devices, we study the effect of a) feature selection and b) the choice of network architecture on variable speaking rates. Four different features are evaluated on multiple configurations of Deep Neural Network (DNN) architectures. The findings show that log Filter-Bank Energies (FBE) outperformed the other acoustic features not only on normal speaking rate but for slow and fast speaking rates as well.

This paper proposes a novel approach that uses deep neural networks for classifying imagined speech, significantly increasing the classification accuracy. The proposed approach employs only the EEG channels over specific areas of the brain for classification, and derives distinct feature vectors from each of those channels. This gives us more data to train a classifier, enabling us to use deep learning approaches. Wavelet and temporal domain features are extracted from each channel. The final class label of each test trial is obtained by applying a majority voting on the classification results of the individual channels considered in the trial. This approach is used for classifying all the 11 prompts in the KaraOne dataset of imagined speech. The proposed architecture and the approach of treating the data have resulted in an average classification accuracy of 57.15%, which is an improvement of around 35% over the state-of-the-art results.

In this work, we propose the use of dropouts as a Bayesian estimator for increasing the generalizability of a deep neural network (DNN) for speech enhancement. By using Monte Carlo (MC) dropout, we show that the DNN performs better enhancement in unseen noise and SNR conditions. The DNN is trained on speech corrupted with Factory2, M109, Babble, Leopard and Volvo noises at SNRs of 0, 5 and 10 dB and tested on speech with white, pink and factory1 noises. Speech samples are obtained from the TIMIT database and noises from NOISEX-92. In another experiment, we train five DNN models separately on speech corrupted with Factory2, M109, Babble, Leopard and Volvo noises, at 0, 5 and 10 dB SNRs. The model precision (estimated using MC dropout) is used as a proxy for squared error to dynamically select the best of the DNN models based on their performance on each frame of test data.

In this paper, I use a homomorphic filter and Deep Neural Network (DNN) for apple trees diseases classification. The homomorphic filter is used as the preprocessing step to enhance appearances of low-level features in an image, which can improve performances of DNN for classification. I experiment on the Plant pathology dataset, which also the Plant Pathology Challenge on Kaggle. The result shows that using a homomorphic filter gets 0.9116 on the accuracy. Keywords-Deep Neural Network, plant disease classification, homomorphic filter, frequency domain image processing.

Review International Paper Image Processing on Medical Image - Medical Images merupakan citra medis yang didapatkan dari pemeriksaan organ / anatomi melalui  alat pemancar gelombang radiografi X-Ray seperti CT-Scan, MRI, dll.

Medical images (Magnetic Resonance Imaging scans) digunakan oleh dokter dan spesialis onkologi untuk menentukan kemungkinan adanya kanker pada paru-paru seorang pasien. Dengan adanya diagnosa terhadap citra medis menggunakan algoritma Deep Neural Network (DNN) akan membantu dokter mengenali kanker pada paru-paru sejak dini. Pendeteksian kanker pada paru-paru secara menyeluruh perlu dilakukan untuk mengetahui keberadaan kanker sejak stadium awal. Sehingga, doktor dapat melakukan pengobatan lebih awal. Oleh karena itu, penelitian ini dilakukan untuk melihat tahapan/stadium mana (0-4) dalam perkembangan kanker yang dapat dideteksi positif kanker menggunakan DNN melalui uji citra kanker paru-paru yang lambat berkembang dengan lebih akurat.

Kontribusi yang diberikan pada penelitian ini yaitu penelitian ini menerapkan arsitektur tambahan pada pengenalan citra dengan Deep Neural Network untuk mengurangi beban komputasi pada Deep Neural Network dan meningkatkan akurasi model klasifikasi yang dilakukan. Arsitektur tambahan pada DNN-nya yaitu penggunaan citra yang telah diklasifikasi dan dilabel sesuai diagnosa oleh spesialis onkologinya (kanker) di rumah sakit medis lokal terlebih dahulu. Klasifikasi ini dilakukan untuk mengurangi beban komputasi DNN dan memastikan pola data training yang digunakan valid dan model pembelajaran pola data training yang dihasilkan bisa dipercayai. Bukan hanya itu, penambahan convolution layer dan max pooling juga dilakukan untuk meningkatkan akurasi klasifikasi.

Ball detection is one of the most important tasks in the context of soccer-playing robots. The ball is a small moving object which can be blurred and occluded in many situations. Several neural network based methods with different architectures are proposed to deal with the ball detection. However, they are either neglecting to consider the computationally low resources of humanoid robots or highly depend on manually-tuned heuristic methods to extract the ball candidates. In this paper, we propose a new ball detection method for low-cost humanoid robots that can detect most soccer balls with a high accuracy rate of up to 97.17%. The proposed method is divided into two steps. First, some coarse regions that may contain a full ball are extracted using an iterative method employing an efficient integral image based feature. Then they are fed to a lightweight convolutional neural network to finalize the bounding box of a ball. We have evaluated the proposed approach using a comprehensive dataset and the experimental results show the efficiency of our method.

We propose a novel architecture that learns an end-to-end mapping function to improve the spatial resolution of the input natural images. The model is unique in forming a nonlinear combination of three traditional interpolation techniques using the convolutional neural network. Another proposed architecture uses a skip connection with nearest neighbor interpolation, achieving almost similar results. The architectures have been carefully designed to ensure that the reconstructed images lie precisely in the manifold of high-resolution images, thereby preserving the high-frequency components with fine details. We have compared with the state of the art and recent deep learning based natural image super-resolution techniques and found that our methods are able to preserve the sharp details in the image, while also obtaining comparable or better PSNR than them. Since our methods use only traditional interpolations and a shallow CNN with less number of smaller filters, the computational c...

This paper provides a comparative performance analysis of both shallow and deep machine learning classifiers for speech recognition task using frame-level phoneme classification. Phoneme recognition is still a fundamental and equally crucial initial step toward automatic speech recognition (ASR) systems. Often conventional classifiers perform exceptionally well on domain-specific ASR systems having a limited set of vocabulary and training data in contrast to deep learning approaches. It is thus imperative to evaluate performance of a system using deep artificial networks in terms of correctly recognizing atomic speech units, i.e., phonemes in this case with conventional state-of-the-art machine learning classifiers. Two deep learning models-DNN and LSTM with multiple configuration architectures by varying the number of layers and the number of neurons in each layer on the OLLO speech corpora along with six shallow machine learning classifiers for Filterbank acoustic features are thoroughly studied. Additionally, features with three and ten frames temporal context are computed and compared with no-context features for different models. The classifier's performance is evaluated in terms of precision, recall, and F1 score for 14 consonants and 10 vowels classes for 10 speakers with 4 different dialects. High classification accuracy of 93% and 95% F1 score is obtained with DNN and LSTM networks respectively on context-dependent features for 3-hidden layers containing 1024 nodes each. SVM surprisingly obtained even a higher classification score of 96.13% and a misclassification error of less than 5% for consonants and 4% for vowels.

We propose a novel architecture that learns an end-to-end mapping function to improve the spatial resolution of the input natural images. The model is unique in forming a nonlinear combination of three traditional interpolation techniques using the convolutional neural network. Another proposed architecture uses a skip connection with nearest neighbor interpolation, achieving almost similar results. The architectures have been carefully designed to ensure that the reconstructed images lie precisely in the manifold of high-resolution images, thereby preserving the high-frequency components with fine details. We have compared with the state of the art and recent deep learning based natural image super-resolution techniques and found that our methods are able to preserve the sharp details in the image, while also obtaining comparable or better PSNR than them. Since our methods use only traditional interpolations and a shallow CNN with less number of smaller filters, the computational cost is kept low. We have reported the results of two proposed architectures on five standard datasets, for an upscale factor of 2. Our methods generalize well in most cases, which is evident from the better results obtained with increasingly complex datasets. For 4-times upscaling, we have designed similar architectures for comparing with other methods.

Adversarial images, inputs designed to produce errors in machine learning systems, are a common way for researchers to test the ability of algorithms to perform tasks such as image classification. "Fooling images" are a common kind of adversarial image, causing miscategorisation errors which can then be used to diagnose problems within an image classification algorithm. Situations where human and computer categorise an image differently, which arise from adversarial images, reveal discrepancies between human image interpretation and that of computers. In this paper, aspects of state of the art machine learning research and relevant artistic projects touching on adversarial image approaches will be contextual-ised in reference to current theories. Harun Farocki's concept of the operative image [1] will be used as a model for understanding the coded and procedural nature of automated image interpretation. Through comparison of current adversarial image methodologies, this paper will consider what this kind of image production reveals about the differences between human and computer visual interpretation .

Human beings live on the time axis and design their daily life accordingly. Mostly they sleep at night and other activities are mostly done at certain times (e.g. breakfast — in the morning). When we consider language, it is known that what is spoken or written maintains and follows certain integrity of meaning. When we talk about machine learning, model training in such sequential datasets is made by using statistical methods such as ARIMA, and SARIMAX, as well as using a neural network. The fundamental idea is that the data at time t is the result of several previous data points. This article explains the theoretical part of RNN — LSTM and includes a tutorial about quick exploratory data analysis of time series dataset and predicting the future power consumption of Germany using LSTM and DNN.

Improving phoneme recognition has attracted the attention of many researchers due to its applications in various fields of speech processing. Recent research achievements show that using deep neural network (DNN) in speech recognition systems significantly improves the performance of these systems. There are two phases in DNN-based phoneme recognition systems including training and testing. Most previous research attempted to improve training phase such as training algorithms, different types of network, network architecture, feature type, etc. But in this study, we focus on test phase which is related to generate phoneme sequence that is also essential to achieve good phoneme recognition accuracy. Past research used Viterbi algorithm on hidden Markov model (HMM) to generate phoneme sequences. We address an important problem associated with this method. To deal with the problem of considering geometric distribution of state duration in HMM, we use real duration probability distribution for each phoneme with the aid of hidden semi-Markov model (HSMM). We also represent each phoneme with only one state to simply use phonemes duration information in HSMM. Furthermore, we investigate the performance of a post-processing method, which corrects the phoneme sequence obtained from the neural network, based on our knowledge about phonemes. The experimental results using the Persian FarsDat corpus show that using extended Viterbi algorithm on HSMM achieves phoneme recognition accuracy improvements of 2.68% and 0.56% over conventional methods using Gaussian mixture model-hidden Markov models (GMM-HMMs) and Viterbi on HMM, respectively. The post-processing method also increases the accuracy compared to before its application.

This paper presents a solution for the problem of making quick changes around ourselves. The objective of "Smart central-service monitoring system for metropolitan city people" is to make an automated and paperless work for the grievances addressed by the public this system enable the automated follow up of the public issues until it is resolved and to provide a quick updates to the citizens for this problems is dropped to the notice of consent person in the department and to the person addressed by them to the government to the elected body.

The image classification is one of the most classical problem of image processing. This research paper about image classification by using deep neural network(DNN) or also known as Deep learning by using framework Tensorflow. Python is used as a programming language because it comes together with Tensorflow framework. Image Classification nowdays is used to narrow the gap between the computer vision and human vision so that the images can be identify by the machine in the same way as human can do. It handle the assigning task for image class. So we are proposing a system called Image Classification using Deep Learning that classifies given images using classifiers such as Neural Network. The system will be built using Python as a programming language and Tensorflow to create neural networks.