Deep Learning: Theory, Architectures and Applications in Speech, Image and Language Processing

Recently, deep learning (DL) computing has become more popular in the machine learning (ML) community. In the field of ML, the most widely used computational approach is DL. It can solve many complex problems, cognitive tasks, and matching problems without any human performance or interface. ML cannot handle large amounts of data and DL can easily handle it. In the last few years, the field of DL has witnessed success in a range of applications. DL outperformed in many application domains, e.g., robotics, bioinformatics, agriculture, cybersecurity, natural language processing (NLP), medical information processing, etc. Despite various reviews on the state of the art in DL, they all concentrated on a single aspect of it, resulting in a general lack of understanding. There is a need to provide a better beginning point for comprehending DL. This paper aims to provide a more comprehensive overview of DL, including current advancements. This paper discusses the importance of DL and introduces DL approaches and networks. It then explains convolutional neural networks (CNNs), the most widely used DL network type and subsequent evolved model starting with LeNET, AlexNet with the Letnet-5, AlexNet, GoogleNet, and ResNet networks, and ending with the High-Resolution network. This paper also discusses the difficulties and solutions to help researchers recognize research gaps for DL applications. 

The emergence of the Internet of Things (IoT) and Artificial Intelligence (AI) applications in many industries is due to recent developments in technology and connectivity. This paper outlines various industry initiatives in healthcare that utilize machine learning techniques. To meet this rising demand, considerable investment is required to develop new medical imaging algorithms, such as those that can be used to diagnose disease diagnostic systems errors, which can yield ambiguous medical treatments. Early disease in imaging is usually predicted by machine learning and deep learning algorithms. Imaging tools use machine learning and deep learning techniques to analyze early disease. Medical imaging is on the cutting edge of deep learning techniques, specifically the application of convolution neural networks. The supervised or unsupervised algorithms are applied to a dataset containing specific instances, and then the predictions are displayed. Machines and deep learning approaches are excellent for data classification and automated decision-making.

The proposed model compares the efficiency of CNN and ResNet50 in the field of digital pathology images. Deep learning methods are widely used in all fields of disease detection, diagnosis, segmentation, and classification. CNN is the widely used image classification algorithm. But it may show less accuracy in case of complex structures like pathology images. Residual Networks are a good choice for pathology image classification because the morphology of digital pathology images is very difficult to distinguish. Colon cancer is one of the common cancers, and it is one of the fatal diseases. If early-stage detection has been done using biopsy results, it will decrease the mortality rate. ResNet50 is selected among the variants as its computational complexity is moderate and provides high accuracy in classification as compared to others. The accuracy metric used here is the training and validation accuracy and loss. The training and validation accuracy of ResNet50 is 89.1% and 90.62%, respectively, whereas the training loss and validation loss are 26.7% and 24.33%, respectively. At the same time, for CNN, the accuracy is 84.82% and 78.12% and the loss is 36.51% and 47.33% .

By detecting lung cancer in advance, doctors can make the right decision to treat patients to ensure that they live long and healthy lives. This research aims to build a CNN model using a pre-trained model and functional API that would classify if a person had lung cancer or not based on a CT scan. This research uses CT scan images as input for the prediction model from the LUNA16 [Luna Nodule Analysis 2016] dataset for experimenting by using ResNet 50 and VGG 16. ResNet50 showed slightly high accuracy on test data compared to VGG16, which is 98%.

Medical image processing takes many steps to capture, process, and convert the images for further analysis. The images are susceptible to distortions due to various factors related to the analysis tools, environment, system-generated faults, and so on. Image enhancement deals with enhancing the quality and resolution of images for accurately analyzing the original information from the images. The primary motivating aspect of research and reconstruction of such high-quality images and their challenges is image super-resolution for image upgrading. This chapter focuses on various image-enhancing strategies in implementing the super-resolution process. In this work, the methodologies of various image-enhancing strategies are explained clearly to provide the parameter selection points, feature comparisons, and performance evaluations that apply to high-resolution image processing. The drawbacks and challenges of each strategy are discussed to investigate the effectiveness of the methodologies. Further research is explored to find hybrid methods on various deep learning architectures to achieve higher accuracy in the field of medical image super-resolution.

Breast cancer is a disease with a high fatality rate each year. It is the most frequent cancer in women and the leading cause of death in women worldwide. The method of machine learning (ML) is an excellent way to categorize data, particularly in the medical industry. It is widely used for decision-making, categorization, and analysis. The main objective of this study is to analyze the performances of different ML algorithms on the WBCD dataset. In this paper, we analysed the performances of different ML algorithms, i.e., XGboost Classifier, KNN, Random Forest, and SVM (Support Vector Machine). Accuracy was used in the study to determine the performance. Experimental result shows that SVMs perform better and are more accurate than KNNs as the amount of training data increases. The SVM produces better results when the main component (PC) value grows and the accuracy rating exceeds the kNN.

Cardiotocography (CTG) is a clinical procedure performed to monitor fetal health by recording uterine contractions and the fetal heart rate continuously. This procedure is carried out mainly in the third trimester of pregnancy. This work aims at proving the significance of upsampling the data using SMOTE (Synthetic Minority Oversampling Technique) in classifying the CTG traces. The project includes the comparison of different Machine Learning approaches, namely, Logistic Regression, Support Vector Machine (SVM), Naïve Bayes, Decision Tree, Random Forest, and Knearest Neighbor (KNN) classifiers on the CTG dataset to classify the records into three classes: normal, suspicious and pathological. The results prove that applying SMOTE increases the performance of the classifiers.

Developing deep neural models for continuous recognition of sign gestures and generation of sign videos from spoken sentences is still challenging and requires much investigation in earlier studies. Although the recent approaches provide plausible solutions for these tasks, they still fail to perform well in handling continuous sentences and visual quality aspects. The recent advancements in deep learning techniques envisioned new milestones in handling such complex tasks and producing impressive results. This paper proposes novel approaches to develop a deep neural framework for recognizing multilingual sign datasets and multimodal sign gestures. In addition to that, the proposed model generates sign gesture videos from spoken sentences. In the first fold, it deals with the sign gesture recognition tasks using a hybrid CNN-LSTM algorithm. The second fold uses the hybrid NMT-GAN techniques to produce highquality sign gesture videos. The proposed model has been evaluated using different quality metrics. We also compared the proposed model performance qualitatively using different benchmark sign language datasets. The proposed model achieves 98% classification accuracy and improved video quality in sign language recognition and video generation tasks. 

Even though the hearing and vocally impaired populace rely entirely on Sign Language (SL) as a way of communication, the majority of the worldwide people are unable to interpret it. This creates a significant language barrier between these two categories. The need for developing Sign Language Recognition (SLR) systems has arisen as a result of the communication breakdown between the deaf-mute and the general populace. This paper proposes a Hybrid Convolutional Recurrent Neural Network-based (H-CRNN) framework for Isolated Indian Sign Language recognition. The proposed framework is divided into two modules: the Feature Extraction module and the Sign Model Recognition module. The Feature Extraction module exploits the Convolutional Neural Network-based framework, and the Model recognition exploits the LSTM/GRU-based framework for Indian sign representation of English Alphabets and numbers. The proposed models are evaluated using a newly created Isolated Sign dataset called ISLAN, the first multi-signer Indian Sign Language representation for English Alphabets and Numbers. The performance evaluation with the other state-o- -the-art neural network models have shown that the proposed H-CRNN model has better accuracy.

Throughout these years, technology has transformed our world and has become an integral part of our everyday lives. This massive digitalization needs to take into consideration the elderly community too. However, the elderly community is usually digitally excluded, but despite these various and diverse difficulties, they remain driven, engaged, and eager to make an effort, to incorporate developing digital technologies into their daily life. Hence, this research implements the functionality of analyzing and determining the emotions of senior citizens by utilizing various natural languages processing-based machine learning techniques like SVM, Random Forest, and Decision Tree. Different algorithms were used, and various parameters were compared to obtain the results. It is seen that SVM gave the best results with an accuracy of 88%. The android application “Bandhu – your forever friend” would be accessible and usable by the elderly community. In addition, this proposed application offers various other features, including learning a new language and listening to the regional songs which would cater to the multicultural requirements in a diverse country like India, recording their audio stories to preserve the ethnic culture and inspire others, and lastly becoming tech-savvy by following easy video tutorials. These features would not only keep them engaged but also tackle their loneliness and isolation to some extent, as all these features were considered after surveying around 100 senior citizens' lifestyles and needs. This app will also make them more digitally independent. All the above-mentioned functionalities have been implemented using the programming language Java and the android application is built in Android Studio. Also, the entire app is in the Hindi language, considering that this language is the most preferred and spoken language in India. 

Before three decades, providing higher education infrastructure for young aspirants in their locality was a challenge for India. With 5164 higher education institutions, including universities, colleges, and stand-alone institutions, India has surpassed the United States as the global leader in educational infrastructure over the last two decades. This work intends to propose an elaborative education ecosystem for sustainable quality education. The secondary data from top global ranking agencies (Times, QS, Webometric, Scimago, and Shanghai Ranking) is deployed to avoid the cost of a worldwide survey for primary data and the execution time. Quality education's quantitative and qualitative parameters are reviewed separately on different scales. The need for the proposed model is evaluated on academic reputation, employer reputation, faculty-student ratio, citations per faculty, international faculty, international students, and infrastructure on the 7-point quality scale. The proposed elaborative model will establish a robust quality education ecosystem on global parameters. The proposed model emphasizes the use of emerging technologies including the Internet of Things (IoT), Artificial Intelligence (AI), and Blockchain (BC), in the education industry. 

Machine learning models, particularly black-box, make powerful decisions and recommendations. However, these models lack transparency and hence cannot be explained directly. The respective decisions need explanation with the help of techniques to gain users' trust and ensure the correct interpretation of a particular recommendation. Nowadays, Knowledge graphs (K-graph) has been recognized as a powerful tool to generate explanations for the predictions or decisions of black-box models. The explainability of the machine learning models enhances transparency between the user and the model. Further, this could result in better decision support systems, improvised recommender systems, and optimal predictive models. Unfortunately, while these black box devices have no detail on the reasons behind their forecasts, they lack clarity. White box structures, on the other hand, will quickly produce interpretations due to their existence. The chapter presents an exhaustive review and step-by-step description for using knowledge graphs in generating explanations for black-box recommender systems, which further helps in generating more persuasive and personalized explanations for the recommended items. We also implement a case study on the MovieLens dataset and WikiData using K-graph to generate accurate explanations.

Retail supermarkets are an essential part of today's economy, and managing them is a tedious task. One of the major problems faced by supermarkets today is to keep track of the items available on the racks. Currently, the track of the product on the shelf is kept by price tag readers, which work on a barcode detection methodology that has to be customized for each store. On the other hand, if barcodes are not present on the price tags, the data is manually fed by the staff of the store, which is really time-consuming. This paper presents a universal pipeline that is based on Optical Character recognition and can be used across all kinds of price tags, and is not dependent on barcodes or any particular type of price tag. This project uses various image-possessing techniques to determine and crop the Area of Interest. It detects the price of the product and the name of the product by filtering the OCR outputs based on the area and dimensions of the bounding boxes of the text detected. Additionally, the presented pipeline is also capable of capturing discounted prices, if any, for the products. It has been tested over price tags of five different types, and the accuracy ranges from 78% to 94.5%. 

Autonomous systems are increasingly being employed in almost every possible field. Their level of autonomy in decision-making is also increasing along with their complexity leading to systems that will soon be making decisions of utmost importance without any human intervention at all or with the least human involvement. It is imperative, therefore, that these machines be designed to be ethically aligned with human values to ensure that they do not inadvertently cause any harm. In this work, an attempt is made to discuss the salient approaches and issues, and challenges in building ethically aligned machines. An approach inspired by traditional Eastern thought and wisdom is also presented. 

Portfolio management is the science of choosing the best investment policies and strategies with the aim of getting maximum returns. Simply, it means managing the assets/stocks of a company, organization, or individual and taking into account the risks, and increasing the profit. This paper proposes portfolio management using a bot leveraging a reinforcement learning environment specifically for cryptocurrencies which are a hot topic in the current world of technology. The reinforcement Learning Environment gives the reward/penalty to the agent, which helps it train itself during the training process and make decisions based on the trial-and-error method. Dense and CNN networks are used for training the agent to taking the decision to either buy, hold or sell the coin. Various technical indicators, like MACD, SMA, etc., are also included in the dataset while making the decisions. The bot is trained on 3-year hourly data of Bitcoin, and results demonstrate that the Dense and CNN network models show a good amount of profit against a starting balance of 1,000, indicating that reinforcement learning environments can be efficacious and can be incorporated into the trading environments.