In recent years, there has been a significant push to implement Deep Neural Networks (DNNs) on edge devices, which requires power and hardware efficient circuits to carry out the intensive matrix-vector multiplication (MVM) operations. This work presents hardware efficient MVM implementation techniques using bit-serial arithmetic and a novel MSB first computation circuit. The proposed designs take advantage of the pre-trained network weight parameters, which are already known in the design stage. Thus, the partial computation results can be pre-computed and stored into look-up tables. Then the MVM results can be computed in a bit-serial manner without using multipliers. The proposed novel circuit implementation for convolution filters and rectified linear activation function used in deep neural networks conducts computation in an MSB-first bit-serial manner. It can predict earlier if the outcomes of filter computations will be negative and subsequently terminate the remaining computations to save power. The benefits of using the proposed MVM implementations techniques are demonstrated by comparing the proposed design with conventional implementation. The proposed circuit is implemented on an FPGA. It shows significant power and performance improvements compared to the conventional designs implemented on the same FPGA.

Human genes often, through alternative splicing of pre-messenger RNAs, produce multiple mRNAs and protein isoforms that may have similar or completely different functions. Identification of splice sites is, therefore, crucial to understand the gene structure and variants of mRNA and protein isoforms produced by the primary RNA transcripts. Although many computational methods have been developed to detect the splice sites in humans, this is still substantially a challenging problem and further improvement of the computational model is still foreseeable. Accordingly, we developed DeepDSSR (deep donor splice site recognizer), a novel deep learning based architecture, for predicting human donor splice sites. The proposed method, built upon publicly available and highly imbalanced benchmark dataset, is comparable with the leading deep learning based methods for detecting human donor splice sites. Performance evaluation metrics show that DeepDSSR outperformed the existing deep learning based methods. Future work will improve the predictive capabilities of our model, and we will build a model for the prediction of acceptor splice sites.

No description available.

Indiana University-Purdue University Indianapolis (IUPUI) / The word “Perception” seems to be intuitive and maybe the most straightforward problem for the human brain because as a child we have been trained to classify images, detect objects, but for computers, it can be a daunting task. Giving intuition and reasoning to a computer which has mere capabilities to accept commands and process those commands is a big challenge. However, recent leaps in hardware development, sophisticated software frameworks, and mathematical techniques have made it a little less daunting if not easy. There are various applications built around to the concept of “Perception”. These applications require substantial computational resources, expensive hardware, and some sophisticated software frameworks. Building an application for perception for the embedded system is an entirely different ballgame. Embedded system is a culmination of hardware, software and peripherals developed for specific tasks with imposed constraints on memory and power. Therefore, the applications developed should keep in mind the memory and power constraints imposed due to the nature of these systems. Before 2012, the problems related to “Perception” such as classification, object detection were solved using algorithms with manually engineered features. However, in recent years, instead of manually engineering the features, these features are learned through learning algorithms. The game-changing architecture of Convolution Neural Networks proposed in 2012 by Alex K [1], provided a tremendous momentum in the direction of pushing Neural networks for perception. This thesis is an attempt to develop a convolution neural network architecture for embedded systems, i.e. an architecture that has a small model size and competitive accuracy. Recreate state-of-the-art architectures using fire module’s concept to reduce the model size of the architecture. The proposed compact models are feasible for deployment on embedded devices such as the Bluebox 2.0. Furthermore, attempts are made to integrate the compact Convolution Neural Network with object detection pipelines.

Managing crowd density has become an immense challenge for public authorities due to population growth and evolving human dynamics. Crowd counting estimates the number of individuals in a given area or scene, making it a practical technique applicable in real-world scenarios such as surveillance and traffic control. It contributes to urban planning, retail analytics, and security systems by providing insights into population dynamics and aiding in anomaly detection. This thesis focuses on implementing and evaluating a selective kernel mechanism in crowd counting. The selective kernel block, introduced in a computer vision research known as the Selective Kernel (SK) Network [1], presents an adapted convolution layer as a substitute for the traditional convolution neural network (CNN) architecture. This adaptation has the potential to enhance object detection and image regression tasks. Building upon the C3 framework [2], the thesis applies the selective kernel mechanism to three state-of-the-art crowd counting designs: ResNet [3], CSRNet [4], and SANet [5], resulting in the creation of SK adaptive models. The evaluation process mainly involves collecting and comparing Mean Absolute Error (MAE) and Mean Squared Error (MSE), as well as crowd statistics and crowd density maps. These evaluations are performed using the ShanghaiTech crowd Part A (random high-density crowd images from the website) and Part B (street views in similar scenes) datasets [6]. In 6 comparisons with two different datasets, SK adaptive models were found to have better prediction results in 4 of them against the original models. In conclusion, the SK block offers several advantages: firstly, it enhances feature extraction performance, especially when pretrained with large datasets; secondly, it improves image regression in more straightforward dataset scenarios. On the downside, its impact is limited or detrimental in sparse datasets. This finding suggests that the selective kernel approach holds promise in supporting and improving crowd counting in the high-density group and street view scenarios, facilitating effective public management. / Att hantera folktäthet har blivit en enorm utmaning för offentliga myndigheter på grund av befolkningsökning och förändrade mänskliga dynamiker. Folkräkning uppskattar antalet individer i ett givet område eller scen, vilket gör det till en praktisk teknik som kan tillämpas i verkliga scenarier som övervakning och trafikstyrning. Genom att erbjuda insikter i befolkningsdynamik och hjälpa till med avvikelsedetektering bidrar folkräkning till stadsplanering, detaljhandelsanalys och säkerhetssystem. Denna avhandling fokuserar på implementeringen och utvärderingen av den selektiva kernelmekanismen inom folksamlingars räkning. Den selektiva kernelblocket, introducerat i en datorseendeforskning känd som Selective Kernel Network [1], presenterar en anpassad faltningsskikt som en ersättning för den traditionella konvolutionsneuralnätverk-arkitekturen. Denna anpassning har potential att förbättra objektdetektion och bildregression. Byggande på C3 - ramverket [2] tillämpar avhandlingen den selektiva kernelmekanismen på tre toppmoderna modeller inom folksamlingars räkning: ResNet [3], CSRNet [4], och SANet [5], vilket resulterar i skapandet av SK-adaptiva modeller. Evalueringen innefattar främst insamling och jämförelse av medelabsolutfel och medelkvadratfel, samt statistik om folksamlingar och densitetskartor. Dessa utvärderingar utförs med hjälp av dataseten ShanghaiTech crowd Part A (slumpmässiga bilder av hög densitet från webbplatsen) och Part B (gatuvyer i liknande scenarier) [6]. Totalt genomförs sex jämförelser med två olika dataset, och SK-adaptiva modeller visar bättre prognosresultat i fyra av dem jämfört med de ursprungliga modellerna. Sammanfattningsvis erbjuder SK-blocket flera fördelar: för det första förbättrar det prestandan för funktionsextrahering, särskilt när det förtränas med stora dataset; för det andra förbättrar det bildregression i enklare dataset-scenarier. Å andra sidan är dess påverkan begränsad eller till och med skadlig i glesa dataset. Generellt sett tyder detta på att den selektiva kärnan har lovande att stödja och förbättra publikräkningen i scenarierna med hög täthet och gatuvy, och därigenom underlätta effektiv offentlig förvaltning.

Yes / One of the common cardiac disorders is a cardiac attack called Myocardial infarction (MI), which occurs due to the blockage of one or more coronary arteries. Timely treatment of MI is important and slight delay results in severe consequences. Electrocardiogram (ECG) is the main diagnostic tool to monitor and reveal the MI signals. The complex nature of MI signals along with noise poses challenges to doctors for accurate and quick diagnosis. Manually studying large amounts of ECG data can be tedious and time-consuming. Therefore, there is a need for methods to automatically analyze the ECG data and make diagnosis. Number of studies has been presented to address MI detection, but most of these methods are computationally expensive and faces the problem of overfitting while dealing real data. In this paper, an effective computer-aided diagnosis (CAD) system is presented to detect MI signals using the convolution neural network (CNN) for urban healthcare in smart cities. Two types of transfer learning techniques are employed to retrain the pre-trained VGG-Net (Fine-tuning and VGG-Net as fixed feature extractor) and obtained two new networks VGG-MI1 and VGG-MI2. In the VGG-MI1 model, the last layer of the VGG-Net model is replaced with a specific layer according to our requirements and various functions are optimized to reduce overfitting. In the VGG-MI2 model, one layer of the VGG-Net model is selected as a feature descriptor of the ECG images to describe it with informative features. Considering the limited availability of dataset, ECG data is augmented which has increased the classification performance. A standard well-known database Physikalisch-Technische Bundesanstalt (PTB) Diagnostic ECG is used for the validation of the proposed framework. It is evident from experimental results that the proposed framework achieves a high accuracy surpasses the existing methods. In terms of accuracy, sensitivity, and specificity; VGG-MI1 achieved 99.02%, 98.76%, and 99.17%, respectively, while VGG-MI2 models achieved an accuracy of 99.22%, a sensitivity of 99.15%, and a specificity of 99.49%. / This project was funded by University of Jeddah, Jeddah, Saudi Arabia (Project number: UJ-02-018-ICGR).

In recent years, the technological space experienced the proliferation of Generative AI models. A prominent type of this model is a language model-based chatbot. The primary function of these models is to generate answers to a question from an extensive database and create a stream of conversation at various levels of complexity. The database of these models encompasses diverse data type of text (e.g., ChatGPT), audio (e.g., PlayHT), or images (e.g., DALLE-2). The intricate process involves neural networks, which undergoes pre-training from the database, building result from architecture neural networks, refined tuning to creating coherent result, probability estimation to produce the correct context result, and generating and refinement as improvement to generated answers. This proposal aims to delve deep into the probability estimation process of the generative AI model. A specific focus is to predict an airborne object's trajectory to create an understanding of how to adapt and adjust robotic limbs and enable them to intercept and capture the thing with some degree of precision.

Falling in indoor home setting can be dangerous for elderly population (in USA and globally), causing hospitalization, long term reduced mobility, disability or even death. Prevention of fall by monitoring different human activities or identifying the aftermath of fall has greater significance for elderly population. This is possible due to the availability and emergence of miniaturized sensors with advanced electronics and data analytics tools. This thesis aims at developing machine learning models to classify fall activities and non-fall activities. In this thesis, two types of neural networks with different parameters were tested for their capability in dealing with such tasks. A publicly available dataset was used to conduct the experiments. The two types of neural network models, convolution and recurrent neural network, were developed and evaluated. Convolution neural network achieved an accuracy of over 95% for classifying fall and non-fall activities. Recurrent neural network provided an accuracy of over 97% accuracy in predicting fall, non-fall and a third category activity (defined in this study as “pre/postcondition”). Both neural network models show high potential for being used in fall prevention and management activity. Moreover, two theoretical designs of fall detection systems were proposed in this thesis based on the developed convolution and recurrent neural networks.

La robotique autonome est employée avec succès dans des environnements industriels contrôlés, où les instructions suivent des plans d’action prédéterminés.La robotique domestique est le challenge des années à venir et comporte un certain nombre de nouvelles difficultés : il faut passer de l'hypothèse d'un monde fermé borné à un monde ouvert. Un robot ne peut plus compter seulement sur ses données capteurs brutes qui ne font qu'indiquer la présence ou l'absence d'objets. Il lui faut aussi comprendre les relations implicites entre les objets de son environnement ainsi que le sens des tâches qu'on lui assigne. Il devra également pouvoir interagir avec des humains et donc partager leur conceptualisation à travers le langage. En effet, chaque langue est une représentation abstraite et compacte du monde qui relie entre eux une multitude de concepts concrets et purement abstraits. Malheureusement, les observations réelles sont plus complexes que nos représentations sémantiques simplifiées. Elles peuvent donc rentrer en contradiction, prix à payer d'une représentation finie d'un monde "infini". Pour répondre à ces difficultés, nous proposons dans cette thèse une architecture globale combinant différentes modalités de représentation d'environnement. Elle permet d'interpréter une représentation physique en la rattachant aux concepts abstraits exprimés en langage naturel. Le système est à double entrée : les données capteurs vont alimenter la modalité de perception tandis que les données textuelles et les interactions avec l'humain seront reliées à la modalité sémantique. La nouveauté de notre approche se situe dans l'introduction d'une modalité intermédiaire basée sur la notion d'instance (réalisation physique de concepts sémantiques). Cela permet notamment de connecter indirectement et sans contradiction les données perceptuelles aux connaissances en langage naturel.Nous présentons dans ce cadre une méthode originale de création d'ontologie orientée vers la description d'objets physiques. Du côté de la perception, nous analysons certaines propriétés des descripteurs image génériques extraits de couches intermédiaires de réseaux de neurones convolués. En particulier, nous montrons leur adéquation à la représentation d'instances ainsi que leur usage dans l'estimation de transformation de similarité. Nous proposons aussi une méthode de rattachement d'instance à une ontologie, alternative aux méthodes de classification classique dans l'hypothèse d'un monde ouvert. Enfin nous illustrons le fonctionnement global de notre modèle par la description de nos processus de gestion de requête utilisateur. / Autonomous robotics is successfully used in controled industrial environments where instructions follow predetermined implementation plans.Domestic robotics is the challenge of years to come and involve several new problematics : we have to move from a closed bounded world to an open one. A robot can no longer only rely on its raw sensor data as they merely show the absence or presence of things. It should also understand why objects are in its environment as well as the meaning of its tasks. Besides, it has to interact with human beings and therefore has to share their conceptualization through natural language. Indeed, each language is in its own an abstract and compact representation of the world which links up variety of concrete and abstract concepts. However, real observations are more complex than our simplified semantical representation. Thus they can come into conflict : this is the price for a finite representation of an "infinite" world.To address those challenges, we propose in this thesis a global architecture bringing together different modalities of environment representation. It allows to relate a physical representation to abstract concepts expressed in natural language. The inputs of our system are two-fold : sensor data feed the perception modality whereas textual information and human interaction are linked to the semantic modality. The novelty of our approach is in the introduction of an intermediate modality based on instances (physical realization of semantic concepts). Among other things, it allows to connect indirectly and without contradiction perceptual data to knowledge in natural langage.We propose in this context an original method to automatically generate an ontology for the description of physical objects. On the perception side, we investigate some properties of image descriptor extracted from intermediate layers of convolutional neural networks. In particular, we show their relevance for instance representation as well as their use for estimation of similarity transformation. We also propose a method to relate instances to our object-oriented ontology which, in the assumption of an open world, can be seen as an alternative to classical classification methods. Finally, the global flow of our system is illustrated through the description of user request management processes.

The master thesis focuses on the selection of a suitable algorithm for the classification of selected network traffic services and its implementation. The theoretical part describes the available classification approaches together with commonly used algorithms and selected network services. The practical part focuses on the preparation and preprocessing of the dataset, selection and optimization of the classification algorithm and verifying the classification capabilities of the algorithm in the various scenarios of the dataset.