Advances in Artificial Intelligence and Applied Cognitive Computing

Predicting if freshmen students will drop out of college or transfer to another is often difficult due to limited and anomalous data. This paper explores using Generative Adversarial Networks (GANs) to learn the general features of student data and uses it to produce predictions with higher accuracy and lower false positive rates than neural networks trained with traditional techniques. Here we examine the differences between a classifier’s latent space when it is trained with a GAN architecture versus traditionally for predicting if a freshman student will leave Marist College within their first year. Our experimental results suggest that GANs are an alternative to training neural models for student dropout/transfer prediction.

Over the last decade, the use of Deep Learning in many applications produced results that are comparable to and in some cases surpassing human expert performance. The application domains include diagnosing diseases, finance, agriculture, search engines, robot vision, and many others. In this paper, we are proposing an architecture that utilizes image/video captioning methods and Natural Language Processing systems to generate a title and a concise abstract for a video. Such a system can potentially be utilized in many application domains, including, the cinema industry, video search engines, security surveillance, video databases/warehouses, data centers, and others. The proposed system functions and operates as followed: it reads a video; representative image frames are identified and selected; the image frames are captioned; NLP is applied to all generated captions together with text summarization; and finally, a title and an abstract are generated for the video. All functions are performed automatically. Preliminary results are provided in this paper using publicly available datasets. This paper is not concerned about the efficiency of the system at the execution time. We hope to be able to address execution efficiency issues in our subsequent publications.

The amount of white blood cells in the blood is of great importance for disease diagnosis. White blood cells include five main classes (eosinophils, lymphocytes, monocytes, neutrophils, basophils), each of which is an important indicator for specific diseases. Deep learning models have been developed to successfully classify the different white blood cell types. The most prominent deep learning models in image classification are deep convolutional neural network (D-CNN) models. A key challenge when solving a problem using deep learning is identifying and setting the hyperparameters for the algorithm. Mostly, these hyperparameters are set manually based on experience. In this study, a new model of deep convolutional neural network is proposed for the classification of four white blood cells types. In this model, the hyperparameters are self-optimized by a genetic algorithm which provides significant improvement in the model. For the verification of the proposed model, four types of white blood cells available from the Kaggle data series were studied. The number of white blood cell images are about 12,000 and are split for training and test sets as 80% and 20%, respectively. When the proposed model was applied to the Kaggle white blood cell data set, the four white blood cell types in the sample data set were classified with high accuracy. The genetic algorithm (GA)–enhanced D-CNN model produced above 93% classification accuracy for the test set demonstrating the success of the proposed enhancement to the D-CNN model with GA. Comparatively, D-CNN models without GA optimization, such as Inception V3 model, produced 84% accuracy, and ResNet-50 model achieved 88% accuracy.

Constituency parse tree is considered the backbone of several Natural Language Processing (NLP) tasks. Deep learning techniques are adopted because they generate parse tree using a dataset without any predefined rules, making them extensible to any language. To capture the semantic meaning, dense words representation technique is necessary. This chapter combines both dense Arabic word representations and deep learning model to generate constituent parse tree. The resultant tree is used in a complete workflow. It contains a web-based application to enable linguists to choose the sentence, generate its constituent, review resultant tree, and edit needed parts. Moreover, the curated output sentence will be used to retrain the model for self-correction. The model is efficient and parallel, resulting in a quick training process.

Using unsupervised Machine Learning (ML) techniques for the discovery of commonalities within a dataset is a well-established approach. However, existing clustering methods require relationships to be represented within the data, making discovery difficult a priori since unlabeled classes are discovered through exploratory clustering. To circumvent exploratory class labeling, we propose a feature-rich, connected structure (i.e., semantic graph), rather than a feature-engineered vectorization, enabling the encoding of maximal information despite class uncertainty. Expanding upon previous tactics discovery work, the authors present a systematic approach using knowledge graph representations and graph embeddings to discover tactics ab initio from data characteristics (DISTRIBUTION STATEMENT A. Approved for public release; distribution unlimited. (OPSEC #4189)).

Today, devices and applications powered by artificial intelligence (AI) include modes of transportation, home appliances, and mobile applications; in short, they are ubiquitous. Many people will have at least heard of AI and perhaps even subdivisions of AI such as Machine Learning (ML) and Deep Learning (DL). Each of these represents an advanced tool of data science explored in this article. First, we briefly review the history of these developments in data science, tracing the history from the first mechanical computer in 1850 to the current state of DL in 2020. Each section overviews some basic definitions, tenets, and current and future developments. We discuss possible future directions of AI—including the transition to Artificial General Intelligence (AGI). Finally, we explore some of the ethical dilemmas posed by such advances and offer a call to data and social scientists to carefully consider the implications of these technologies.

Magnetic resonance image (MRI) is widely applied to the brain tumor diagnosis and treatment. Approximately 35 million MRIs are performed annually worldwide in recent years. Manual segmentation and extraction of the tumor area from MRIs are time-consuming. The emergence of deep-learning algorithms offers new opportunities to automate the medical images processing and analysis with high accuracy. In the study, we built deep-learning models for brain tumor segmentation. The MRI files were first preprocessed through the procedures including reorientation, de-noising, bias-correcting, skull stripping, and co-registration. Then, the two deep-learning algorithms, DeepMedic and 3D U-Net, were used for tumor segmentation model construction. Different from the sequential DeepMedic model, 3D U-net has an encoding and decoding patch that allows shortcut connections from layers with equal resolution in the encoding path to the layers in the decoding path and can provide the high-resolution features. The dice coefficient (DICE), a most commonly used metric for validating medical volume segmentations, was adopted for performance evaluation. Our DeepMedic model achieved DICE of 0.802, whereas 3D U-Net achieved 0.876 for overall segmentation. Moreover, we built a linear regression model using shape features including size and surface area of different segmented tumor tissue sections from the results of 3D U-Net model along with clinical data of age and gross total resection status for patient survival prediction. Compared to using the clinical data alone, we found that combining shape features improved the prediction of overall survival by 7%. We further increased the overall survival prediction accuracy by an additional 9% by replacing the shape features that were not significantly correlated with survival with some selected texture. Our work provided models for automated brain tumor segmentation and patient survival prediction.

Explored building a generative model with the combination of variational autoencoder (VAE) and generative adversarial network (GAN) that offers better result when the agent interacts with the environment. Our agent model can train the unsupervised environment and increase the imaging quality. Moreover, it provides better control option and produce better accuracy results as compared to traditional systems. An experiment performs on the car racing based upon the designed agent model and features can extracted effectively that help to produce reliable information. With the combination of variational autoencoder and generative adversarial network, it provides better feature extraction to gather relevant data and solves the complexity. With the combination of VAE and GAN, elevated-level visual abstraction can be made effectively.

Chemistry dynamics simulation is widely used in quantitative structure activity relationship QSAR, virtual screening, protein structure prediction, quantum chemistry, materials design, and property prediction, etc. This chapter explores the idea of integrating Long Short-Term Memory (LSTM) with chemistry dynamics simulations to enhance the performance of the simulation and improve its usability for research and education. The idea is successfully used to predict the location, energy, and Hessian of atoms in a H2O reaction system. The results demonstrate that the artificial neural network–based memory model successfully learns the desired features associated with the atomic trajectory and rapidly generates predictions that are in excellent agreement with the results from chemistry dynamics simulations. The accuracy of the prediction is better than expected.

In Entity Resolution (ER), more and more unstructured records impose challenge to the traditional similarity-based approaches, since existing similarity metrics are designed for structured records. Now that similarity is hard to measure for unstructured records, can we do pairwise matching without similarity measure? To answer this question, this research leverages deep learning’s artificial intelligence to learn the underlying record matched pattern, rather than measuring records similarity first and then making linking decision based on the similarity measure. In the representation part, token order information is taken into account in word embedding, and not considered in Bag-of-Words (Count and TF-IDF); in the model part, multilayer perceptron (MLP), convolutional neural network (CNN), and long short-term memory (LSTM) are examined. Our experiments on both synthetic data and real-world data demonstrate that, surprisingly, the simplest representation (Count) and the simplest model (MLP) together get the best results both in effectiveness and efficiency. An F-measure as high as 1.00 in the pairwise matching task shows potential for further applying deep learning in other ER tasks like blocking.

In recent years, the development of robots has been carried out for making human life more convenient and more comfortable along with the development of artificial intelligence. It is necessary for the robot to recognize the surrounding environment. However, in the surrounding environment there are objects other than real objects such as illustrations and paintings. When recognizing an image showing an illustration image with the current object recognition system which learned using real-object images, the recognition rate is very low (about 65%). In this research, we aim to recognize both illustration images and real-object images, and we verified whether the pseudo illustrated image which processed contour processing and the color reduction processing to the real image is effective for the recognition of the illustrated image.

Skin cancer is the most commonly diagnosed cancer in the United States with over a million cases being detected each year. Fortunately, early detection provides high odds of recovery. The traditional method of detection involves clinical screening, which is prone to false positives, followed by an invasive biopsy. While this provides for a high rate of detection, it is intrusive and costly. Artificial Intelligence for medical image analysis has proved effective in assisting in the diagnosis of many medical maladies, yet fine variations in the appearance of skin lesions has made applications to skin cancer detection difficult. We report that a deep convolutional neural network (CNN) trained over clinically labeled images (pixels) can accurately assist in the diagnosis of early-stage skin cancers. Specifically, we analyze skin lesions using CNN and evaluate its performance on seven dermatologist-certified clinical image types: Actinic keratoses and intraepithelial carcinoma (Bowen’s disease), basal cell carcinoma, benign keratosis-like lesions (solar lentigines, seborrheic keratoses, and lichen-planus-like keratoses), dermatofibroma, melanoma, melanocytic nevi, and vascular lesions (angiomas, angiokeratomas, pyogenic granulomas, and hemorrhage). The model provides significantly high levels of average accuracy, specificity, and sensitivity across these types.

A requirement for a significant amount of training as well as the exploration of potentially expensive or safety-critical states limits the applicability of reinforcement learning for real-world robotics. One potential solution is given by pretraining models in simulations before transferring them to the real world. In this chapter, we investigate the concept of domain randomization to train robust agents in simulation to control an industrial robot. We examine the effects of different degrees of randomization with respect to the transferability to the real world. In addition, we use attention maps to gain insights into the agents’ decision-making processes. We find that attention maps enable a qualitative assessment for the data-efficiency of a pretrained agent when transferred to the real-world setup.

In this study, we use motion recognition to recognize unseen and unlabeled movement patterns, which are widely used and challenging in machine learning. Motion recognition tackles some of the emerging challenges in computer vision problems, such as analyzing actions in a surveillance video where there is a lack of sufficient training data. Motion recognition also plays a pivotal role in human action and behavior recognition. In this paper, we propose a novel action and motion recognition method using zero-shot learning. We overcome a limitation of machine learning by recognizing unseen and unlabeled classes in the field of human action recognition. In order to evaluate the effectiveness of the proposed solution, we use a dataset available from the UCI machine learning repository. This dataset enables us to apply zero-shot learning to human motion and action recognition. Our results verify that the proposed method outperforms state-of-the-art algorithms.

The objective of this project is to determine if data mining techniques may be applied to small data sets to forecast population growth of isolated populations of organisms, with a useful level of accuracy. The Isle Royale moose population was chosen to be the basis of this study, because of the quality of available data and substantial previous work studying the Isle Royale wolf/moose population dynamics.

This study exploits previously demonstrated properties (i.e., sensitivity to spatial extent and intensity of local image contrasts) of the quantization error in the output of a Self-Organizing Map (SOM-QE). Here, the SOM-QE is applied to double-color-staining-based cell viability data in 96 image simulations. The results from this study show that, as expected, SOM-QE consistently and in only a few seconds detects fine regular spatial increase in relative amounts of RED or GREEN pixel staining across the test images, reflecting small, systematic increase or decrease in the percentage of theoretical cell viability below a critical threshold. While such small changes may carry clinical significance, they are almost impossible to detect by human vision. Moreover, here we demonstrate an expected sensitivity of the SOM-QE to differences in the relative physical luminance (Y) of the colors, which translates into a RED–GREEN color selectivity. Across differences in relative luminance, the SOM-QE exhibits consistently greater sensitivity to the smallest spatial increase in RED image pixels compared with smallest increases of the same spatial magnitude in GREEN image pixels. Further selective color contrast studies on simulations of biological imaging data will allow generating increasingly larger benchmark datasets and, ultimately, unravel the full potential of fast, economic, and unprecedentedly precise predictive imaging data analysis based on SOM-QE.

Learning systems that adapt to learner capabilities, needs, and preferences have been shown to improve learning outcomes. However, creating systems that can interpret learner state within the context of a dynamic learning environment is costly and often tied to the specific requirements of the learning environment. We overview a new approach for monitoring and assessing system context and learner state that is not specific to a particular domain. The process is designed to transform diverse, continuous, and multichannel streams of heterogeneous system data into a consistent, discrete, and learner-centric interpretation of the situation. Key steps in the process include discretizing the data stream into “events” and then marking some events as “keyframes” that identify important steps or changes in the learning state. This keyframing process provides a compact representation for use by learner-adaptive processes (including assessment and tailoring) and simplifies the challenges of using machine learning as a mechanism for adaptation.

Streaming data on power usage delivered through the Advanced Metering Infrastructure (AMI) inherent the concept drift problem, in which the shape of the data changes over time. This phenomenon causes performance degradation in the processing of AMI data using deep learning models. In order to overcome this, updates of deep learning (DL) models using dataflow in real-time should be performed. Although there have been many studies tried to handle the issue so far, the problem with existing methodologies is that they have not fully considered the factors that affect the training efficiency of online learning in an environment where concept drift exists.In this paper, adaptive online learning techniques are proposed to solve the issue. This technique first determines batch size and epoch size by considering data integration latency and unstableness of the current DL model. Then it adjusts the epoch according to concept drift existence of current training input batch. By applying this technique, we were able to effectively reduce the load on learning time series power data using the DL model, while at the same time showing better performance in forecasting with newly incoming data.

Currently, with the high rate of generation of new information, it is important the traceability of its evolution. This paper studies techniques that allow analyzing the evolution of the knowledge, starting with analyzing the capabilities of the techniques to identify the patterns that represent the common information in datasets. From the “patterns,” the evolution of their characteristics over time is analyzed. The paper considers the next techniques for the problem of tracking the traceability of the patterns: LDA (Latent Dirichlet allocation), Birch (Balanced Iterative Reducing and Clustering using Hierarchies), LAMDA (Learning Algorithm for Multivariate Data Analysis), and K-means. They are used both for the initial task of grouping the data, as well as, to analyze the characteristics of the patterns, and the relevance of them in the patterns through their evolution (traceability). This paper uses different types of data sources of educational contents, and with these datasets, the topological models to describe the “patterns” generated from the grouping of the analyzed data, and their dynamics (evolution over time), are Studied (traceability). For the evaluation, the paper considers three metrics: Calinski–Harabasz Index, Davies–Bouldin Index, and Silhouette Score.

This paper presents an interactive approach for exploratory data analysis of StarCraft: BroodWar replay data. Our approach encodes the basic facts of StarCraft units into statements in first-order logic (FOL). The information from replay data is also encoded into FOL statements. This database of FOL facts, in conjunction with a logic programming language, enables a human user to easily query and augment the database with new facts. This approach encourages an interactive workflow to aid in the analysis of replay data. We present examples that demonstrate the utility of the approach.

The World Health Organization declared a global pandemic of Covid-19 on March 11, 2020. Upon originally animal-hosted virus presented inside human bodies, over the time, human-adaptive transmissibility can be induced, and once a huge number of humans get infected by the virus, its transmutability can be further enhanced to lead to a global pandemic due to nowadays ever faster flow of humans from place to place and much rapid transportation network around the world. With the advent of high-throughput next-generation sequencing technology, virus genomes can be sequenced promptly, hence paved a way for us to develop deep learning approaches to predict human-adaptive transmutability and transmissibility. We propose a Convolutional Neural Network to impute the data and build an innovative Deep Learning model for predicting adaptive interactions with the host cells. Moreover, we propose to construct a Heterogeneous Information Network integrating multilayer data in combination with the lineage information to infer host–cell-specific and lineage-specific networks. The hypothesis underlying our approach is that viral mutation and the subsequent adaptive interaction networks drive the evolution of human adaptive transmutability. Hence, based on the novel representation embedded with heterogeneous information, a Simulated Annealing method is proposed to search for the global optimization to study the information fusion of communities at the interaction networks. Finally, using the fusion information, we propose a novel ranking-based clustering for master interaction detection. The results can lead to better understand how viral transmutability affects the interaction networks to induce human adaptive transmissibility and, better understand human adaptive transmissibility of the disease, as well as advance the identification of potential drug targets for effective treatments.

Human activity recognition is one of the major research problems in computer science and it has a wide range of applications including healthcare applications. The goal of this research project is to design and apply machine learning methods that can be used to analyze publicly available datasets collected by wearable sensors of elderly and identify different activities such as setting and walking. Feature selection and ranking algorithms were used to select the most relevant feature and reduce the number of collected and used data from sensors. Several classification algorithms were used in this chapter for activity recognition and several experiments were conducted to compare between these different algorithms using different number of features based on the ranking of features. The experimental results show a high accuracy in recognizing activities can be achieved using a fewer number of features. This can help in providing the medical services required for elderly based on the detected activity using a small number of collected features from sensors. The experimental results show that random forest algorithm gave the highest accuracy compared to the other algorithms using only three features of sensors data.

How might machine learning be employed to help promote the ethos of democratic citizenship in popular society? This chapter uses machine learning tools to better understand the role political culture plays in determining levels of political participation among citizens spread across our democratic republic. Although many studies have utilized political culture to explore rates of political participation along class, economic, gender, and other such cleavage lines, many scholars focusing on this framework often neglect the importance of race. In what regional political culture are we more likely to find higher levels of political participation among White and Black Americans? We investigate this question. We also seek to determine whether Black Americans have their own political culture that transcends previous understandings of regional political subcultures? We show that Elazar’s classification of regional political subcultures applies mostly to White Americans, and that political culture among Blacks do not adhere to geographical lines or region. Machine learning tools are deployed to sift through massive data and uncover patterns and structures embedded within it. The machine learning tools here were used to test model specification and will be used to improve its performance to better predict political participation among Americans. The information may be used by policymakers and concerned citizens in their effort to increase civic participation and combat political apathy.

In this paper we present SentiTel, a fine-grained opinion mining dataset that is human annotated for the task of targeted aspect-based sentiment analysis (TABSA). SentiTel contains Twitter reviews about three telecoms in Uganda posted in the period between February 2019 and September 2019. The reviews in the dataset have a code-mix of English and Luganda a language that is commonly spoken in Uganda. The dataset in this paper consists of 5973 human annotated reviews with the target entity which is the target telecom, aspect and sentiment towards the aspect of the target telecom. Each review contains at least one target telecom. Two models are trained for the TABSA task that is random forest which is the baseline model and BERT. The best results are obtained using BERT with an Area Under the ROC Curve (AUC) of 0.950 and 0.965 on aspect category detection and sentiment classification respectively. The results show that even though tweets are written without the intention of writing a formal review, they are rich in information and can be used for fine-grained opinion mining. Finally, the results show that fine-tuning the pre-trained BERT model on a downstream task generates better results compared to the baseline models.

Recently, knowledge graph-based personalized recommendation systems have been proposed to exploit rich structured information. In particular, knowledge graph reasoning via knowledge embedding is increasingly introduced to preserve semantic information in the knowledge graph. However, due to the large dimensional representations of the millions of entities and relations, knowledge graph reasoning requires excessive computational resources. Moreover, selecting meaningful latent features in the knowledge graph between users and items is a challenging issue. In this paper, to impose efficient and accurate personalized recommendations, we propose a path-based recommendation using Q learning. Our model also offers a good explanation for recommendations. Experimental results show that the proposed method provides higher accuracy, compared to a uniform random walk based algorithm.

The past several years have seen a strong push toward using Deep Learning systems–Neural Networks with multiple hidden layers. Deep Learning is now used in many machine learning applications and provides leading performance on numerous benchmark tasks. However, this increase in performance requires very large datasets for training. From a practitioner prospective, the model that performs best in benchmark tasks may be too data intensive to be adapted to practical application. We describe a behavior recognition problem that was solved using a sequence-based Deep Learning system and then reimplemented using a more knowledge-driven sequence matching approach due to data constraints. We contrast the two approaches and the data required to achieve sufficient performance and flexibility.

For a robot to converse naturally with a human, it must be able to accurately gauge the emotional state of the person. Techniques for estimating the emotions of a person from facial expressions, intonation and speech content have been proposed. This chapter presents a technique for judging the emotion of a person from EEGs by SVM using pseudo extended data. The accuracy of emotion judgment from EEG features using at random pseudo extended data was 25.0% and using proposed extension technique was 30.0%. However, performance accuracy remains low, and continued development is required through further development of methods for both reducing noise mixed with EEGs.

Safety is an important issue at large gatherings of people such as at religious gatherings or sporting events. Therefore, it is important to control crowds and identify in advance when dangerous situations may arise. Simulations play an important role in predicting how people in crowds will react to one another and their environment. Simulations often rely on a priori models of human behavior to predict crowd behavior. We combine genetic algorithms and neural networks to learn how people behave in crowds to avoid assumptions used in a priori models. We examine learning in two important regimes: structured crowds, where individuals are moving in a specified direction, as in the Islamic Hajj or Hindu Kumbh Mela, and unstructured crowds, such as town squares and train stations. In this preliminary work, we are most concerned with questions of trainability. In order to provide sufficient data and control qualitative features of our crowd data, we begin use generated data based on elaborations of wildlife flocking models in NetLogo. We compared performance on structured and unstructured crowds by predicting a series of next locations. The results showed that we are able to predict crowd motion, but error rates for individuals grow as time passes; however, the structured crowd gave more reliable results than the unstructured crowd.

In this chapter, we present a hybrid application of genetic algorithms to a problem of optimizing a car trajectory in a car race setting. The aim is to find a curve traversing the race track, or a trajectory, that minimizes the total distance, and hope that it allows a car to finish the race faster. For this, we start by mapping and segmenting the track and then setting non-uniform key frames based on significant intervals where the track is almost straight and on long curves going in a single direction. The trajectory’s value is anchored in the key frame points, insuring an overall smoothness. Then on each interval, we apply the genetic algorithm to compute the trajectory, and piece it together at the end. We compare the results with a procedurally computed trajectory and with trajectories obtained by the genetic algorithm with uniform key framing.

Although an Artificial Neural Network (ANN) is a biology-mimicking system, it is built from components designed/fabricated for use in conventional computing, and it is created by experts trained in conventional computing; all of them are using the classic computing paradigm. As von Neumann in his classic “First Draft” warned, because the data transfer time is neglected in the model he used, using a “too fast processor” vitiates the procedure; furthermore, that using his paradigm for imitating neuronal operations is unsound. This means that at least doubly unsound to apply his paradigm to describe scaling ANNs. The common experience shows that making actively cooperating and communicating computing systems, using segregated single processors, has severe performance limitations; and that fact cannot be explained using his classic paradigm. The achievable payload computing performance of those systems sensitively depends on their workload type, and this effect is only poorly known. The type of the workload that the Artificial Intelligence (AI)-based systems generate leads to an exceptionally low payload computational performance. Unfortunately, the initial successes of demo systems that comprise only a few “neurons” and solve simple tasks are misleading: the scaling of processor-based ANN systems is strongly non-linear. The chapter discusses some major limiting factors that affect their performance. It points out that for building biology-mimicking large systems, it is inevitable to perform drastic changes in the present computing paradigm; namely, instead of neglecting the transfer time, a proper method to consider it shall be developed. The temporal behavior enables us to comprehend the technical implementation of computing components and architectures.

Pruning is a legitimate method for reducing the size of a neural network to fit in low SWaP hardware, but the networks must be trained and pruned offline. We propose an algorithm, Artificial Neurogenesis (ANG), that grows rather than prunes the network and enables neural networks to be trained and executed in low SWaP embedded hardware. ANG accomplishes this by using the training data to determine critical connections between layers before the actual training takes place. Our experiments use a modified LeNet-5 as a baseline neural network that achieves a test accuracy of 98.74% using a total of 61,160 weights. An ANG grown network achieves a test accuracy of 98.80% with only 21,211 weights.

Digital images possess rich features that are highly correlated among neighboring pixels and highly redundant bits. Thus, the study of encryption algorithms over images is a subject of study to determine encryption quality. In this chapter, we study the applicability of neural networks in learning to use secret keys to protect information (e.g., images) from other neural networks. We implement a well-known adversarial neural network architecture that is designed to learn to establish its own encryption method to protect itself from an attacker that is also a neural network. Our preliminary results suggest that this type of neural architecture is capable of securing communications in the midst of a neural-based attacker using only 16-bit blocks.

Nowadays, by the integration of renewable energy resources, the expenditures pertaining to energy production are diminished with respect to the vast advancements in their science and technology. Considering the integration of renewable energy resources in power system operation, particularly in the generation section, poses grave difficulties in market clearance and calculation of unit commitment problem in order to achieve the optimum objectives. The inclusion of uncertainty in a large-scale non-linear non-convex non-smooth optimization problem enhances the complexity of the problem. To solve such a sophisticated problem, some exact or heuristic methods have been proposed so far. In this chapter, a 10-generator test system is taken into account for the simulation of the proposed method, and the modified social spider algorithm is used to solve the unit commitment problem considering the impact of the presence of wind farms. The constraints related to the spinning reserve requirements are also incorporated to maintain system security. Finally, economic load dispatch is performed to find the optimum generation level of all committed units and obtain operational costs. The problem procedure is repeated regardless of the presence of wind units to compare the results and to assess the effectiveness of the proposed approach.

Climate change is causing longer forest fire seasons and more fires in places where they traditionally were rare. Accordingly, the US Forest Service’s annual budget devoted to wildfires jumped from 16% to over 50% between 1995 and 2015 and exceeded $2.4 billion in 2017. Allocating and supplying the correct amount of personnel and equipment to a fire as quickly as possible is vital in reducing acreage destroyed, costs, and saving lives. In this chapter, we explore the possibility of predicting the number of personnel needed to effectively fight a wildfire, in an aim to develop a forecasting system in the future. On the IRWIN test dataset, our model obtained a coefficient of determination (R2) 0.61 and mean absolute error 1.4 personnel. Thorough analysis and discussion are also provided in the chapter.

Credit risk prediction is a task that continues to be studied extensively for various crediting schemes. Machine learning has become a viable option for developing credit risk prediction models for uncertain cases involving a considerable number of instances. In particular, the framework of Bayesian networks is a very suitable option for managing uncertainty. Although there have been several studies on the use of Bayesian networks in credit risk prediction, there is scarce literature about their application to cases from developing economy contexts. In this chapter, we exhaustively apply and evaluate several Bayesian network models for credit risk prediction based on cases from a Ugandan financial institution. Credit risk prediction quality from some Bayesian network models is satisfactory and compares with prediction quality from other state-of-the-art machine learning methods. Evaluation results show that one Bayesian network model learned through a global optimization hill climbing method always leads to the highest prediction quality so far on Ugandan credit contracts.

Artificial feelings and emotions are beginning to play an increasingly important role as mechanisms for facilitating learning in intelligent systems. What is presented here is the theory and architecture for an artificial nervous/limbic system for artificial intelligence entities. Here we borrow from the military concept of operations management and start with a modification of the DoD Observe, Orient, Decide and Act (OODA) loop. We add a machine learning component and adapt this for processing and execution of artificial emotions within an AI cognitive system. Our concept, the Observe, Orient, Decide, Act, and Learn (OODAL) loop makes use of locus of control methodologies to determine, during the observe and orient phases, whether the situation constitutes external or internal controls, which will affect the possible decisions, emotions, and actions available to the artificial entity (e.g., robot). We present an adaptation of the partial differential equations that govern human systems, adapted for voltage/current regulation rather than blood/nervous system regulation in humans. Given human trial and error learning, we incorporate a Q-learning component to the system that allows the AI entity to learn from experience whether its emotions and decisions were of benefit or problematic.

In multi-agent systems, agents’ abilities are often used to classify a system as either homogeneous or heterogeneous. In the context of multi-agent reinforcement learning (MARL) systems, the agents can also be homogeneous or heterogeneous in their strategies. In this work, we explore instances where agents with homogeneous capabilities must collaborate to achieve a common goal in a predator–prey pursuit task. We show that results from homogeneous and heterogeneous strategies associated with learning differ substantially from agents with fixed strategies that are analytically defined. Given that our agents are homogeneous in capability, here we explore the impact of homogeneous and heterogeneous strategies in a MARL paradigm.

Historically, research shows analysis, characterization, and classification of complex heterogeneous non-linear systems; and interactions have been difficult to accurately understand and effectively model. Synonymously, exponential growth of Internet of Things (IoT), cyber physical systems, and the litter of current accidental and unscrupulous cyber events portray an ever-challenging security environment wrought with complexity, ambiguity, and non-linearity, thus providing significant incentive to industry and academia toward advanced, predictive solutions to reduce persistent global threats. Recent advances in artificial intelligence (AI) and information theoretic methods (ITM) are benefitting disciplines struggling with learning from rapidly increasing data volume, velocity, and complexity. Research shows axiomatic design (AD) providing design and datum disambiguation for complex systems utilizing information content reduction. Therefore, we propose a transdisciplinary AD, AI/ML, ITM approach combining axiomatic design with advanced, novel, and adaptive machine-based learning techniques. We show how to significantly reduce risks and complexity by improving cyber system adaptiveness, enhancing cyber system learning, and increasing cyber system prediction and insight potential where today context is sorely lacking. We provide an approach for deeper contextual understanding of disjointed cyber events by improving knowledge density (KD) (how much we know about a given event) and knowledge fidelity (KF) (how well do we know) ultimately improving decision mitigation quality and autonomy. We improve classification and understanding of cyber data and reduce system non-linearity and cyber threat risk, thereby increasing efficiency by reducing labor and system costs, and “peace of mind.”

Wave Function Collapse (WFC) is an iterative constraint solving algorithm that performs texture synthesis on input samples to generate procedural outputs. The two commonly used WFC implementations are the Overlapping WFC (OWFC) and Tiling WFC (TWFC) implementations. OWFC incurs a performance cost in identifying constraints whereas TWFC receives pre-determined constraints in the form of meta-data. Pre-determining the meta-data, however, is a non-trivial process and requires substantial design time before the execution of the algorithm. The proposed Markov WFC (MkWFC) implementation aims to reduce the time involved during the meta-data design stage while maintaining the performance benefits of the TWFC implementation. This is achieved by using Markov random fields to determine constraints from a set of input samples and introduces a pre-processing step to the TWFC implementation. By automating constraint identification, the MkWFC implementation reduces both the meta-data generation time as well as the scope for human error. This paper compares TWFC against MkWFC on a set of 6 different procedural image generation problems using 3 sizes for inputs and outputs in a total of 720 trials. When comparing the MkWFC implementation against TWFC, there is an increase in the number of identified and used constraints from input and output respectively, which increases with image size. The performance of TWFC and MkWFC was compared by measuring their respective execution times for all trials. MkWFC was able to identify over 1.5 times more constraints for a 5 × 5 tiled image in 1.03 ± 0.09 ms(α = 0.05) and almost 3 times more constraints in 25 × 25 tiled image in 28.19 ± 2.58 ms(α = 0.05). This is substantially faster than the TWFC methodology where these constraints have to be manually identified and entered into the meta-data file. The automated meta-data generation and nominal increase in execution time allows for MkWFC to scale where TWFC cannot.

In steel industry, automatic defects inspection and classification is of great importance to improve the quality. This chapter proposes and develops parallel algorithms using CUDA to improve the required computing time to detect defects in surface steel strips. The algorithm divides steel images into non-overlapped region of interest (ROI) and employs the summed area table to improve the required time to extract statistical features per (block) ROI. The computation time of the proposed parallel algorithm excels the sequential one. Support vector machine classifier has been used to classify patches, scratches, and scale defects. The experimental results indicate significant improvements and 1.6 speed up.

In this paper we propose a method to approximate softmax layer for computer vision applications, especially on the devices with limited hardware (HW) resources, such as mobile or edge platforms. In this paper we showed that using a max-normalization in the inverse way as x i ∗ = m a x ( x ) − x i $$x^*_i = max(x)-x_i$$ together with the substitution of e x by its reciprocal 1∕e x allows to obtain an efficient formula for inference on the devices with constrained resources. To validate our method we have conducted experiments with human segmentation model over dataset with 1,000 images. It is shown that even ultra-low quantization down to 2-bit is applicable, maintaining a negligibly small accuracy loss (0.06% for 2-bit quantization). The required size of look-up-table (LUT) is also small (3 to 8 Bytes only).

The paper deals with the implementation of similarity-based approach decisions which are considered in detail (specific areas and cases in real life). The role of similarity, descriptive similarity, and numerical evaluation is comprehensively examined. The author pointed out that descriptive similarity is the relevant approach for equation of the statements. The paper introduces the transport situation similarity results in town of Ostroh (Ukraine) and certain Estonian towns, which is of great use for Ostroh public transportation claims. The author indicated that it is important to transform textual information into formulas. Artificial intelligence application needs of two level are under appreciate importance. To justify these decisions suitable algebraic formulas of system theory as well as experts’ logical steps are used.

In this study, our focus is on the design of mission scheduling techniques capable of working in dynamic environments with unmanned aerial vehicles, to determine effective mission schedules in real time. The effectiveness of mission schedules for unmanned aerial vehicles is measured using a surveillance value metric, which incorporates information about the amount and usefulness of information obtained from surveilling targets. We design a set of dynamic heuristic techniques, which are compared and evaluated based on their ability to maximize surveillance value in a wide range of scenarios generated by a randomized model. We consider two comparison heuristics, three value-based heuristics, and a metaheuristic that intelligently switches between the best value-based heuristics. The novel metaheuristic is shown to find effective solutions that are the best on average as all other techniques that we evaluate in all scenarios that we consider.

In recent years, location-based advertising (LBA) can deliver advertisements to customers in targeted locations and provide product and service information of their local businesses. Thus, understanding customers’ needs and considerations is essential for the popularization of LBA services. The purpose of this study is to explore the key factors that customers would concern about while they decide to turn on Bluetooth on their mobile devices to receive LBA services using fuzzy-AHP in conjunction with DEMATEL method. The analysis results in this study indicate that “personal location privacy,” “get information about saving money right time and right place,” and “personal preference privacy” are the top three consideration factors.

In this study, we present how to divide chromosome data sets into scalable hierarchical clusters. Diagnosis applications of chromosomal abnormalities to identify genetic diseases are mostly implemented with semi-automatic methods, which do not provide high-throughput and fast characterizations of patient specimens. However, selecting and managing specific features of chromosome data sets require dynamic and scalable data models. Here, we adapt the variations in sets to feature units as an effective tool for our dynamic/automated and scalable approach. This method enables geneticists to implement karyotyping techniques easily into an efficient identification of hereditary disorders or characterization of novel chromosomal alterations related with genetic diseases in a trusted/scalable manner. Furthermore, we explore physical limits of available application-specific integrated circuits (ASICs) and on-board computers (OBCs) to extract novel features of chromosomes via spectral analytics for real-time diagnosis and to overcome the bottlenecks of the computational complexity, as well.

Descriptions of situations immediately preceding the outbreak of a military conflict, specifically eve of military invasion are under consideration. Situation descriptions are treated as sets of relevant statements. The descriptions of various situations highlight similar claims. Descriptions containing identifiable statements form associations. Developments and situations observed and treated through descriptions that consist of statements. One form of similarity of situations or developments observed as descriptive similarity and its results through numerical evaluation. Based on of descriptive similarity, the eve of military invasion, involving states attack to another state, are investigated. The calculations made express the irrelevance of the corresponding descriptions from public available sources. Means provided to quantify the general similarity of descriptions of situations constituting an association. Such estimates provide a first insight into what and how similar situations may develop in the context of this association.

Music recommendation systems have become an important part of the user-centric online music listening experience. However, current automated systems often are not tuned for exploiting the full diversity of a song catalogue, and consequently, discovering new music requires considerable user effort. Another issue is current implementations generally require significant artist metadata, user listening history, or a combination of the two, to generate relevant recommendations. To address the problems with traditional recommendation systems, we propose to represent artist-to-artist relationships as both simple multigraphs and more complicated multidimensional networks. Using data gathered from the MusicBrainz open music encyclopedia, we demonstrate our artist-based networks are capable of producing more diverse and relevant artist recommendations.

In this paper, we propose a low-cost video analytics system analyzing multiple video streams efficiently under limited resource. The objective of our proposed system is to find the best configuration decision of frame sampling rate for multiple video streams in order to minimize the accuracy degradation in the shared limited resource, utilizing the velocity features extracted from video context in low cost. To evaluate the proposed algorithm, we use a subset of videos from VIRAT dataset. The results show that our video analytics system outperforms the existing video analytics systems on resource-accuracy trade-offs and reduces the high profiling cost of them.

Video surveillance system with object re-identification is cited as a challenge to address to enhance the safety and convenience of citizens. The system consists of a combination of complex tasks requiring a lot of computing workload. With these characteristics, efforts have continued to accelerate the system. Existing systems did not benefit from the service latency perspective to make good use of heterogeneous accelerated edge computing system. In this paper, the goal is to accelerate the system used in smart cities on limited heterogeneous edge servers, and the scheduling planning method considering them is proposed. We first identify the computational volume-based execution time model of the heterogeneous accelerators. Then, we propose a scheduling plan that distributes this task graph to resources. Finally, the planning method proposed in this paper is experimentally compared with the previous GPU-FPGA allocation scheme. We compare it to the previously proposed method, and show that queue latency can be reduced, with showing robustness to the deadline violation rate.

Video game designers express a need for a tool to create a human-like Non-Player Character (NPC) that makes contextually correct decisions against inputs performed at unforeseen times. This tool must be able to integrate with the designer’s current game engine, such as the Unreal Engine. To make an NPC more human-like, we use psychology theories, specifically, emotions in decision making, such as the ideas proposed by A. Damasio. This approach allows an NPC’s psychology to connect directly to the underlying intrinsic motivations that drive players. This paper presents the “Artificial Psychosocial Framework” (APF) and defines a new class of NPC known as an “Affective NPC” (ANPC). APF utilizes an ANPC’s personality, perceptions, and actions to generate emotions, which influences the social relationship they have with other ANPCs. Additionally, APF considers the generation of emotions as an ambivalent hierarchy and classifies the representation of a personality and social relation. APF can work in tandem with existing AI techniques, such as Fuzzy Logic, Behavior Trees, Utility Theory, and Goal-Oriented Action Planning, to provide an interface for developers to build emotion-centric gameplay. APF is a library for designers to use as a plug-in to a familiar game engine, such as the Unreal Engine. On the macro-level, APF uses the “Cognitive Appraisal Theory” (CAT) of emotions. On the micro-level, psychological models are chained together to determine the emotional state and social relation. The personality of an ANPC is represented by the Big Five: “Openness, Consciousness, Extroversion, Agreeableness, and Neuroticism,” also called the OCEAN model. The emotions of an ANPC are described by the “Ortony, Clore, and Collins” (OCC) model. The social relation between ANPCs uses the “Liking, Dominance, Solidarity, and Familiarity” (LDSF) model. This paper demonstrates APF by simulating a scenario from the “Dennis the Menace” setting, where each Dennis is an ANPC with a different personality mapped to a Myers-Briggs type indicator. The results demonstrate how a different personality profile will cause varying emotional states, responses, and actions for the two simulated ANPCs in the same simulated scenario.

The NNEF (Neural Network Exchange Format) is a de facto standard file format for the neural network description and exchange. In this work, we represent a simple implementation of the NNEF execution system, which is similar to the programming language interpreters. While the original NNEF file format focused on the data exchange format itself, our prototype implementation shows that the contents of an NNEF file can be directly executed by the underlying computing systems.

With the development of the web, a large amount of product reviews might accumulate within short time, and the problem of information overload is getting serious. Previous opinion analysis studies on online reviews have focused on finding out the positive and negative opinions on the function of commodities. But the comments, which indicate particular emotion of the reviewer, were rarely discussed. For hedonic products such as music that could invoke users’ feelings, the emotional rules including such as “happiness” and “sadness” mining from the comments can further complement the traditional rules, such as “singer is good” and “cannot sing high pitch.” Taking the example of joy emotion analysis, this study proposed a system structure to classify the emotional feeling of pop music online reviews. A prototype system was built. The experiment reports the satisfactory result, with F1 73.45%.

This paper presents a low-cost design with half-, single-, and double precision on a MultiLayer Perceptron (MLP) neural network. It contains one input layer, one hidden layer, and one output layer. And in each layer, multiple sigmoid neurons are implemented to train and recognize handwritten digits. The weights and biases of this network are trained using the Modified National Institute of Standards and Technology (MNIST) handwritten digit data in Python using the stochastic gradient descent method. By running the design network on Matlab, the accuracy of the three designs can reach over 93% accuracy. This work basically offers a preliminary result of a hardware demonstration on field-programmable gate array (FPGA).

The electrocardiogram (ECG), a recording of the electrical activity of the heart, is commonly used for cardiac analysis, but lack of abnormal ECG signal data restricts the development of high quality automatic auxiliary diagnosis. In this paper, we introduce an LSTM and GAN based ECG abnormal signal generator to alleviate the issue. By training with a small set of real abnormal signals, the proposed generator can learn and produce high quality fake abnormal signals. The fake signals are then combined with real signals to train abnormal ECG classifiers. We show that our method can significantly improve the ability of classifiers in recognizing the uncommon case with a low proportion in the database.

This paper presents a hardware architecture, IoT-Edge-Server, of a diverse embedded system including a wide variety of applications such as smart city, smart building, or smart agricultural farm. First of all, we improve computation time by integrating the idea of edge computing on Raspberry Pi and CPU, which processes different algorithms. Second, the hardware processors are connected to a server that can manipulate the entire system and also possess storage capacity to save the system’s important data and log files. Specifically, the hardware computes data from (1) a non-standardized Bluetooth Low Energy (BLE) mesh system and (2) a surveillance system. The BLE mesh system has one master and three slave devices, while the surveillance system has a passive infrared (PIR) sensor and a camera to detect motion. Experimental results prove that using the phenomena of edge computing demonstrates an improvement in computation speed and data privacy.Our vision is thus to create a system as a case study, capable of sensing the surrounding environment, and more importantly, directing different types of sensor data to the optimal place, in terms of computing devices, for analysis and making decisions autonomous at the proximity of the network edge to improve data privacy, latency, and bandwidth usage.

This chapter proposes a demonstration of edge detection on field-programmable gate array (FPGA), enabling to detect the edge of 320 × 240 size of images at 1302 frames per second (fps). The future work is an integrated system on chip (SoC) with a low-cost bus architecture, a security engine, and an image/video processing data path including OV7670 camera and VGA-interfaced display. The end goal will be a demonstration and simulation on self-driving vehicle to detect obstacles at the network edge. The design of many intellectual properties (IPs) of the SoC has been made publicly available to serve research and teaching courses at University of Houston-Clear Lake (UHCL), as well as to bring together researchers in other universities with interests in integrated circuit design, robotics, and FPGA prototyping.

This chapter presents a novel approach to revolutionize STEM education with the computational thinking (CT) and Active teaching model and its implementation for skill knowledge learning (Act -STEM). We examine the effectiveness of using computational thinking on students’ understanding and competence in STEM discipline learning. The work is built upon two successful pilot projects that demonstrated a broad impact on student learning outcomes.

Studies show that at a young age, children unexceptionally exhibit intense interests in depicting and replicating objects from their external world. The most common expression is through drawing and painting. However, it is not an easy task for most of the population. Artistic expression, especially visual art, has long been proven to be remarkably beneficial for STEM (science, technology, engineering and math) education (Tytler, How art is drawing students to STEM, Australian Council for Educational Research (ACER) conference 2016 - improving STEM learning: What will it take?, 2016; Tyler-Wood et al. J Technol Teach Educ 18(2):341–363, 2010). How to effectively impart such important skill knowledge to the majority of students who lack innate artistic talent is an important and difficult task for researchers. We have developed an effective approach with software solution to help students who lack visual art talent to effectively master the skill knowledge of realistic drawing and painting of objects in the world, including human, animal, plant, scene, architecture, machinery, design, etc. The preliminary result shows our approach is very promising.

Daily peak load forecasting (DPLF) plays a crucial role in unit commitment, security analysis, and scheduling of outages and fuel supplies in smart grid applications. Recently, various artificial intelligence-based DPLF models have been proposed for accurate electric load forecasting using sufficient datasets. However, if the available data are not sufficient for training, it is not easy to build an accurate DPLF model. Herein, we propose a novel DPLF scheme that can perform DPLF well even when the dataset for training is not sufficient. We first configured various input variables by preprocessing time and weather data, as well as the historical electric load. Then, we performed a time-series cross-validation to consider as many training datasets as possible. Simultaneously, we generated new input variables relevant to the daily peak load by using principal component analysis and factor analysis and considered them to build our DPLF model. To evaluate the performance of our approach, we employed it for day-ahead peak load forecasting and verified that our scheme can achieve better prediction performance than traditional methods.

Various services are being embodied for smart city using IT technologies. For instance, face recognition technology can be used for effectively search for missing persons or track criminals. The key to face recognition is to detect landmarks that are the main features of faces. Many studies have been done to detect them accurately and quickly, but there is still much to be done. Deep learning methods for facial landmark detection mainly have used convolutional neural networks. Recently, a new attempt to go beyond CNNs is being tried by considering relationships among all pixels in the images to improve missing long-range context problem in CNNs. In this chapter, we propose a scheme for improving the performance of facial landmark detection based on attention and show its performance through various experiments.

Intelligent CCTV-based surveillance is becoming an essential element in smart cities. Despite the recent explosion of CCTV installed for security purposes, its monitoring still depends on people. Person re-identification is a technique to find an image in disjoint camera views that contains the previously detected pedestrian. Conventional methods for person re-identification used the similarity based on hand-crafted features and their performance heavily relies on lighting or camera angle. In recent years, deep learning-based methods have shown good performance in person re-identification. However, deep learning-based models using two input images have a limitation that they cannot detect similarities and differences between images simultaneously. In this chapter, we propose a model that calculates similarities and differences between images simultaneously by extracting features from the input of three images and reconstructing the extracted feature map.

Smart grids can optimize their energy management by analyzing data collected from all processes of power utilization in smart cities. Typical smart grids consist of diverse systems such as energy management system and renewable energy system. In order to use such systems efficiently, accurate load forecasting should be carried out. However, if there are many anomalies in the data used to construct the predictive model, the accuracy of the prediction will inevitably decrease. Many statistical methods proposed for anomaly detection have had difficulty in reflecting seasonality. Hence, in this chapter, we propose VAE (Variational AutoEncoder)-based scheme for accurate anomaly detection. We construct diverse artificial neural network-based load forecasting models using different combinations of anomaly detection and data interpolation, and then compare their performance. Experimental results show that using VAE-based anomaly detection with a random forest-based data interpolation shows the best performance.

Healthcare practices include collecting all kinds of patient data which would help the doctor correctly diagnose the health condition of the patient. This data could be simple symptoms observed by the subject, initial diagnosis by a physician or a detailed test result from a lab. Thus, far this data is only utilized for analysis by a doctor who then ascertains the disease using his/her personal medical expertise. The Artificial Intelligence has been used with Naive Bayes classification and Random Forest classification algorithm to classify disease datasets of heart disease, to check whether the patient is affected by that disease or not. A performance analysis of the disease data for both algorithms is calculated and compared. The results of the simulations show the effectiveness of the classification techniques on a dataset, as well as the nature and complexity of the data set used.

Next to skin cancer, the most common cancer disease found in women is the breast cancer and which very rarely can found in men. This type of cancer will develop on the upper part of the breast on the lobules which causes lobular carcinoma or in the milk ducts which causes ductal carcinoma which will be usually in form of tumor cells which is visible in mammogram X-ray image or feel like lumps. The cancer becomes deadly when the cell starts growing and it spreads and invade the skin tissues which are around the breast. About 15% of the cancer deaths are because of breast cancer as per the survey of World Health Organization (WHO). At the same time, all the lumps cannot be as deadly as the breast cancer. So identifying the disease and classifying its type in the early stage will reduce the impact. In this research, we propose a hybrid deep learning–based approach for detecting and classify the breast cancer cells in the early stage. The algorithm used in this proposed work is a hybrid approach, which is a combination of Convolutional Neural Networks (CNN) and Random Forest (RF) algorithms that are the most popular deep learning algorithms for detection of breast cancer. The mammogram dataset for this research was collected from various online dataset repositories. Experimental results show that the proposed method gains an overall accuracy of 98.6% with sensitivity and specificity of 96.5% and 98% respectively. While compared with the other state-of-the-art methods, the superior performance of the proposed hybrid technique is visible. The overall system was implemented in MATLAB 2018b software with deep learning toolbox.

In the fast-moving world, obesity has become a major health issue to the human beings. BMI defines the obesity when it is greater than 30 kg/m2. Obesity leads to many diseases like high cholesterol, liver failure, knee problems, diabetes, and sometimes cancer. When the patient eats healthy food, the obesity can be controlled. The obesity problem can be addressed when there is a system that monitors the food consumed by the patient automatically and gives the suggestion periodically to the patient in treatment of obesity. Many of the people find difficulty in monitoring their food intake periodically, due to less knowledge in nutrition and self-control. In this chapter, identification of food type is made and estimation of calorie is done using MLP and proposes the results. Single food item types were considered previously, but here mixed food item types are considered. Region of Interest (ROI) is used to identify the mixed food item type. The next step includes feature extraction process. The extracted feature image is fed into MLP classification to classify the food image. The volume of the food is used to calculate the calories present in the food. The implementation is processed in MATLAB with 1000 fruit images containing 6 food classes with good accuracy. The automatic dietary control is made available for the diabetic patients.

Diabetes mellitus (DM) is one of the main medical issues far and wide causing national financial weight and low personal satisfaction. People with diabetes have an extended possibility of glaucoma. The hurt brought about by glaucoma is irreversible. This can occur if abnormal vein growth, which can occur because of diabetic retinopathy, is the significant consequence of diabetic illness is diabetic retinopathy (DR), it will hinder the characteristic misuse of the eye which impacts the retina of diabetic individuals, and the basic period of diabetic retinopathy can prompt changeless vision misfortune. The early discovery and observation of diabetic retinopathy are critical to forestall it or for compelling treatment, yet the issue related to early identification of diabetic retinopathy (DR) is minor changes on retinal fundus picture, and it incorporates hemorrhages, exudates, red sore, cotton fleece spots, drusen, and so forth. The early location or screening of changes on the retinal picture is exceptionally testing and tedious for ophthalmologists, as the size and shading changes are at first coordinated with neighborhood veins I retinal picture. So the glaucoma is one of the most unsafe visual maladies, keeps on influencing and weight a huge area of our populace. Accordingly, it is basic to distinguish glaucoma at the beginning. The proposed frameworks have focused on the parameter cup to plate proportion (CDR) for identification of glaucoma that might be the best methodology for building proficient, vigorous, and precise computerized framework for glaucoma diagnosis, and this strategy advocates the utilization of half and half methodology of manual element making with profound learning. It holds the guarantee of improving the precision of glaucoma conclusion through the robotized systems.

Road traffic accidents are a “global tragedy” that generates unpredictable chunks of data having heterogeneity. To avoid this heterogeneous tragedy, we need to fraternize and categorize the datasets. This can be done with the help of clustering and association rule mining techniques. As the trend of accidents is increasing throughout the year, agglomerative hierarchical clustering approach is proposed for time series big data for trend analysis. This clustering approach segments the time sequence data into different clusters after normalizing the discrete time sequence data. Agglomerative hierarchical clustering takes the objects with similar properties and groups them together to form the group of clusters. The paradigmatic time sequence (PTS) data for each cluster with the help of dynamic time warping (DTW) is identified that calculates the closest time sequence. The PTS analyzes various zone details and forms a cluster to report the data. This approach is more useful and optimal than the traditional statistical techniques.

A successful implementation of DR (demand response) always depends on proper policy and their empower technologies. This paper proposed an intelligent multiagent system to idealize the residential demand response in distributed network. In our model, the primary stakeholders (smart homes and retailers) are demonstrated as a multifunctional intelligent agents. Home agents (HAs) are able to predict and schedule the energy load. Both HAs and RAs are modelled to predict the real-time pricing. We used LSTM model (artificial neural networks) to predict the electricity load and energy price. Simulation results present significant reduction in electricity payments.

K-nearest neighbors (K-NN) is an effective lazy and instance-based learner. It is natural to scale the K-NN to big data. In this paper, we propose to conduct spectral clustering through t-distributed stochastic neighboring embedding (t-SNE) and then apply k-nearest neighbor to improve its performance and accuracy on big data.

Electricity is one of the key role players to build an economy. Electricity consumption and generation can affect the overall policy of the country. This opens an area for some intelligent system that can provide future insights. Intelligent management for electric power consumption requires future electricity power consumption prediction with less error. These predictions provide insights for making further decisions to smooth line the policy and help to grow economy of the country. Future prediction can be categorized into three categories namely (1) long-term, (2) short-term, and (3) mid-term predictions. For our study, we consider mid-term electricity consumption prediction. Dataset is provided by Korea Electric power supply to get insights for metropolitan city like Seoul. Dataset is in time series so we require to analyze dataset with statistical and machine learning models that can support time series dataset. This study provides experimental results from the proposed models. Our proposed models for provided dataset are ARIMA and LSTM, which look promising as RMSE for training is 0.14 and 0.20 RMSE for testing.

Electricity is playing an important factor to drive the economy of the nation. Every country is trying to find fuel resources alternative to gasoline. Electricity is the promising resource because of low carbon footprints as compared to other fuel resources. Right now, biggest electricity consumers are households and industries. Forecasting the need of the respective sectors, governments can decide the future direction. This can result in better planning. As the second phase of our project, we have tested LST with Bi-LSTM to check the overall performance of the neural network model. Dataset is provided by Korea Electric power supply to get insights for metropolitan city like Seoul. Dataset is in time series so we require to analyze dataset with time distributed machine learning models that can support time series dataset. This study provides experimental results from the proposed models. Our model shows RMSE scores of 0.15 on training and 0.19 for testing with tuning hyperparameters of the model to optimum level.

Blockchain and Bitcoin have both become a very important topic in academia as well as in the economic world. This is due to the main characteristics of this system: transparency, anonymity, security, and low transactional costs. This paper studies the blockchain network, with special attention to Bitcoin. Its uses, advantages, and challenges are all presented, with special focus in instability and volatility, trust and security, cyberattacks, and hacking as well as illegal activities. Within the conclusion, the main future research trends are also analyzed.

The exponential growth of multimedia content on the Internet has led to the generation of search engines and user-specific recommender systems. However, not all practices carried out through these systems are legal, a condition that will depend on the object of the search or recommendation and the country where such practice is carried out, presenting a special prosecution as far as certain aspects of intellectual property are concerned. In this paper, we analyze the legality of search engines and recommender systems in audiovisual content platforms from the point of view of copyright and trademarks mainly, distinguishing the factors to be taken into account when incorporating into their system the titles of works that do not have in their catalogue.

This extended abstract shows an overview of the development of a model-driven decision support system (DSS) to ensure efficient management of workload priorities in organizations. The DSS has been built through discrete-event simulation (DES) using guidelines dictated by Beer’s Viable System Model (VSM). The prioritization policy considers four key factors: customer importance, nature of task, value of task and window of opportunity. They can be appropriately combined by the use of the DSS according to the set of indicators defined by the organization.

In this short research paper, we conceptually propose a valuation model that is suitable for assessing R&D projects and managing uncertainty in pharmaceutical laboratories. Taking this into consideration, this model would allow these organizations to take better strategic decisions that will affect the pipeline of clinical trials (per phase) and the portfolio of innovative drugs. More specifically, our valuation methodology would help decision makers to identify if and when to promote and abandon clinical trials and new drug developments. To this end, we adopt a real options valuation approach, which is combined with fuzzy techniques and simulation. Our proposal incorporates some additional features in relation to the previous literature that aim to make our model more adaptable to deal with real-world uncertainties in the development of new drugs.

Establishment of locations is a very important task in the Named Entity Recognition technology. The given paper describes a contrastive study of the prepositional constructions with a spatial meaning in the Russian and Chinese languages. Basing on the principles of cognitive semantics, a spatial use of the Russian preposition ‘na’ (meaning ‘on’ in English) is investigated in combination with NP in the prepositional case in the Russian language compared with the frame construction ‘在NP上’ (zài NP shàng) in Chinese. For semantic representation of these constructions, the pictures of image schemata are used, i.e. the mental structures that reflect the spatial relationships of ‘trajector’ and ‘landmark’. The analysis of the markers compared in Russian and Chinese shows that although the underlying schemata they reflect are largely the same, there is no full translational equivalence between these markers. This is due not only to the fact that for one of the image schemata expressed by the Chinese construction in the Russian language, the preposition ‘nad’ (‘over’) is used but also by a number of other factors: competition of alternative schemas for the same referent spatial configuration and preference by the Russian language of the variant encoded by another preposition (lampa pod potolkom ‘a lamp under the ceiling’ > lampa pod potolkom ‘a lamp on the ceiling’; dyrka v podoshve ‘a hole in the sole’ > dyrka na podoshve ‘a hole of the sole’), by a stronger influence of the topological type of a landmark object and the nature of the trajector movement on the choice of the preposition in the Russian language.

The research and development results are presented that implement semantic clustering methods based on vector models, extraction of conceptual structures in a multilingual mode, and realization of semantic Internet navigation within the specified domains. Automatic construction of high-quality sustainable topic hierarchies is designed to solve a wide range of artificial intelligence tasks, including the automatic construction of reviews of mass media sources, identifying misinformation, extremist activities, advertisement targeting, and forecasting prospective directions of science and technology development. As the Internet grows around the world, users write comments in different languages. Multilingual methods of tuning neural networks training have been developed to analyze text data in different languages.

Modern education requires the formation of thinking in complexness, which allows the cognition and creation of post-non-classical holistic texts of the culture of mankind and to synchronize at various levels of mutual assistance, up to global. The presented mental model of educational environments is based on the description of complexness relationships at two levels of self-implementation of life forms with consciousness: levels of energy-matter-information model of the world (EMI mw) and the model of the world of culture.

The success of global optimization rests on the ability for a compatible metaheuristic to approximate global search. Particle swarm optimization (PSO) is one of such heuristics, with the ideal PSO application being one that promotes swarm diversity while incorporating the global progress of the swarm in its performance. In this paper, the authors introduce an adaptive tribal topology within PSO to improve global coverage. Diversity of the swarm population was dynamically managed through an evaluation of a swarm fitness parameter, which takes into account the relative performance a swarm member and its assigned tribe has on finding better objective evaluations. The fitness function simultaneously promotes the breeding of exemplars, and the elimination of swarm members stucks in local minima. The model was evaluated on a series of benchmark problems with unique and diverse search spaces, and the results demonstrate that performance relied on the distribution and scale of the local minima present.

Previous work has focused on the overall theory of systems-level thinking for artificial intelligent entities in order to understand how to facilitate and manage interactions between artificial intelligent system and humans or other systems. This includes the ability to predict and produce behaviors consistent with the overall mission (duties) of the AI system, how to control the behaviors, and the types of control mechanisms required for self-regulation within an AI entity. Here we advance that work to look at the overall systems AI thinking process (SATP) and the architecture design of self-regulating AI systems-level processes. The overall purpose here is to lay out the initial design and discussion of concepts to create an AI entity capable of systems-level thought and processing.

In the era of big data, outlier detection has become an important task for many applications, such as the network intrusion detection system. Data streams are a unique type of big data, which recently has gained a lot of attention from researchers. Nevertheless, there are challenges in applying traditional outlier detection algorithms for data streams. One of the well-known algorithms of outlier detection is Local Outlier Factor (LOF). The issue with LOF is that it needs to store the whole dataset with its distances’ results in memory. In addition, it needs to start from the beginning and recalculate all processes if any change happens in the dataset, such as inserting a new data point. The Genetic-based Incremental Local Outlier Factor (GILOF) has addressed these issues and shown significant results. In this paper, we further improved the GILOF performance in data streams by proposing a new calculation method for LOF. The experiment’s results showed that our new method of calculation has led to better accuracy performance for various real-world datasets.

In an era where machine learning algorithms are widely used in order to improve the performance of network intrusion detection system, the complexity and big volume of data available in the network are also on the rise. The cyber networks frequently encounter high-dimensional, unreliable, and redundant data that are often too large to process. An efficient feature selection can therefore remove the redundant and irrelevant attributes and select relevant attributes that can significantly improve the overall system performance. This research provides a variance fractal dimension feature selection method to explore the significant features of cyber security attack dataset. A complexity analysis was done to find out the cognitive discriminative features of UNSW-NB15 dataset. A performance comparison is also provided using our proposed methodology for an artificial neural network, and a comparative analysis was also done that shows the proposed method helps improve the detection performance in network system. The resultant discriminative features not only consume less resource but also speed up the training and testing process while maintaining good detection rates.

Outlier detection is getting significant attention in the research field of big data. Detecting the outlier is important in various applications such as communication, finance, fraud detection, and network intrusion detection. Data streams posed new challenges to the existing algorithms of outlier detection. Local Outlier Factor (LOF) is one of the most appropriate techniques used in the density-based method to determine the outlier. However, it faces some difficulties regarding data streams. First, LOF processes the data all at once, which is not suitable for data streams. Another issue appears when a new data point arrives. All the data points need to be recalculated again significantly. Therefore, it affects the execution time. A new algorithm is proposed in this research paper called Grid Partition-based Local Outlier Factor (GP-LOF). GP-LOF uses a grid for the LOF with a sliding window to detect outliers. The outcome of experiments with the proposed algorithm demonstrates the effectiveness in both performance accuracy and execution time in several real-world datasets compared to the state-of-the-art DILOF algorithm.

It has always been a challenge to make meaning out of unstructured data. In the field of network intrusion detection, the availability of structured, labeled datasets is limited. The need of unsupervised learning from unlabeled datasets is of vital significance, yet there is little breakthrough achieved in the research community. Most approaches adhere to techniques that are over-exhaustive in terms of resources and do not yield satisfactory results; hence, human analysts must re-examine all the events for intrusion attempts. This study makes an effort to find an approach of making sense out of unstructured, unlabeled data, in a way that helps the human analysts to disregard a major portion of the network dataset that contains regular traffic and isolates the finite time-windows that have been subjected to potential attacks, utilizing the concepts of cognitive science, complexity analysis, and statistical higher-order feature learning. In this research, use statistical higher-order features from network flows to classify the network traffic into flows containing normal traffic and flows subject to attacks, using unsupervised k-means clustering and variance fractal dimension trajectory-based complexity analysis. We validate our algorithm on the UNSW dataset and compared our results with traditional unsupervised clustering. The proposed model was able to detect errors with the accuracy of 87.27%.

The long-term goal of artificial intelligence (AI) is to provide machines the capabilities to learn, think, and reason like humans. To achieve these long-term goals, it is necessary to introduce human cognitive-like abilities into AI systems to create truly self-adaptive artificially intelligent systems. This marriage of human cognitive skills with “machines” creates hybrid systems that have characteristics of both. The question becomes which human cognitive model is appropriate for hybrid artificial intelligent systems. The purpose of this paper is to discuss the development of cognitive models to be infused into a modern radar system to create a cognitive radar system (CRS). The notion of a hybrid artificially intelligent system can be divided into two main categories: (a) human-in-the-loop systems with hybrid augmented intelligence requiring human-AI communication/collaboration and (b) a cognitive computing-based AI in which a fully cognitive model is infused into the machine to allow fully autonomous operation. Here, we discuss the first type, human-in-the-loop cognitive radar systems that provide intelligent decision support and analysis for radar systems. The design of hybrid artificial intelligence methods and algorithms is presented with applications to improvement to modern radar systems, utilizing a Rasmussen Cognition Model (RCM), which we feel is appropriate for a hybrid cognitive system utilized to create a cognitive radar system (CRS).

A fierce search is called for a reliable, non-intrusive, and real-time capable method for assessing a person’s experienced cognitive load. Software systems capable of adapting their complexity to the mental demand of their users would be beneficial in a variety of domains. The only disclosed algorithm that seems to reliably detect cognitive load in pupillometry signals—the index of pupillary activity (IPA)—has not yet been sufficiently validated. We take a first step in validating the IPA by applying it to a working memory experiment with finely granulated levels of difficulty, and comparing the results to traditional pupillometry metrics analyzed in cognitive research. Our findings confirm the significant positive correlation between task difficulty and IPA the authors stated.

In recent years, chaos theory has been applied to neuroscience to help understand the human brain. Researchers have suggested that the brain is a dynamical system which behaves chaotically. This paper proposes the introduction of chaos into an artificial neural network (ANN) by using chaotic neurons in the network’s inner layer. A chaotic neuron uses two activation functions (AF): the sigmoid function and the logistic map function. The chaotic neuron generates a unique activation value to send to each neuron in the following layer. The model was tested by solving the XOR problem and showed significant improvements over a typical ANN.

Recently, the aviation industries showed interest in transferring their aircraft models to cyber-physical system (CPS)-based models. However, and as it is well known, CPS introduces security threats to the physical components of the system. Distributed Denial-of-Service (DDOS) attack is one of the significant security threats to the availability of the communication network, due to making the network resources unavailable. In the presence of DDOS attack, servers are overwhelmed with so many requests which increase the network traffic that causes packet loss. Therefore, this paper proposes an approach to defending the communication network system in the aviation cyber-physical system (ACPS) from a DDOS attack using a nonlinear autoregressive exogenous (NARX) model. NARX network is used to predict packets that were dropped due to a DDOS attack. The real-time simulation results showed that the network system in ACPS was successfully defended from the DDOS attack because the ACPS maintained the expected normal performance during the DDOS attack.

This paper’s focus is on using simulated annealing (SA) to improve basic particle swarm optimization (PSO) by embedding SA in the part of PSO that is responsible for a particle’s inertial force, creating a new hybrid evolutionary algorithm. The paper studies the effect of this addition using two benchmark functions in a continuous solution space as well as the effect of this addition on PSO for the Travelling Salesperson Problem (TSP), which has a discrete solution. Finally, the novel hybrid algorithm is compared against basic PSO to establish a ground-level performance evaluation.

This paper introduces an information extraction (IE) architecture to process unstructured data that are typically generated within non-formal contexts such as the ones that are exchanged during multiple forms of technical meetings. Information extraction and knowledge management are providing practical value in all kinds of industries and boosting R&D activities specifically. Our project sees a clear opportunity to leverage and expand this capability to process also “informal sources” of business/technical content such as the one that is exchanged along informal channels during, for instance, physical or web meetings. Given these observations, we have defined a software architecture designed on the concept of expandability; a prototype has also been developed capable of showing the peculiarities and the critical points.