Samira Sorayassa and Majid Ahmadi, Department of Electrical and Computer Engineering,University of Windsor, Windsor, Ontario, N9B3P4, Canada
Tangent Hyperbolic (Tanh) has been used as a preferred activation function in implementing a multi-layer neural network. The differentiability of this function makes it suitable for derivative-based learning algorithm such as error back propagation technique. In this paper two different memory-based techniques for accurate approximation and digital implementation of the Tanh function using Look Up Table (LUT) and Range Addressable Look Up Table (RALUT) are given. A thorough comparative study of the two techniques in terms of their hardware resource usage on FPGA and their accuracies are explained. The schematic of the synthesized design for special cased are given as an example.
Tanh Activation function, Tanh Implementation on FPGA, Approximation methods, Lookup Tables (LUT) Range Addressable Lookup Tables (RALUT).
Christophe Guyeux, University of Burgundy, France
For a number of years now, the regional fire department centers have been recording their interventions numerically. Such databases are under-utilized and are mainly used for statistical and management purposes. However, such a history of interventions can be very useful, if used in conjunction with artificial intelligence algorithms, for predictive purposes. Such work has recently been done in France through a series of articles investigating the various aspects of the problem, and has been put into production at the Doubs center. The objective of this review is to take stock of all the work that has been done so far, to list the successes and the stumbling blocks, and to draw up a roadmap on this theme for the years to come.
Sukhamay Kundu, Dept of Computer Science and Engineering Louisiana State University, Baton Rouge, LA 70803, USA
The regression-line for a set of data-points pi = (xi,yi),1 = i = N, lacks the rotation-property in the sense that if each pi is rotated by an angle ? around the origin then the regression-line does not rotate by the same angle ? except for the special case when all pi’s are collinear. This makes the regression-line unsuitable as a linear model of a set of data points for applications in data mining and machine learning. We present an alternative linear model that has the rotation property. In many ways, the new model is also more appealing intuitively as we show with examples. The computation of the new linear model takes the same O(N) time as that for the regression-line.
perpendicular distance, regression-line, rotation property, application to data mining.
Ali Anaissi, Widad Alyassine, Basem Suleiman and Kenneth D’souza, School of Computer Science, The University of Sydney, Australia
Structural Health Monitoring aims to utilise sensor data to assess the integrity of structures. Machine learning is opening up the possibility for more accurate and informative metrics to be determined by leveraging the large volumes of data available in modern times. An unfortunate limitation to these advancements is the fact that these models typically only use data from the structure being modeled, and these data sets are typically limited, which in turn limits the predictive power of the models built on these datasets. Transfer learning is a subfield of machine learning that aims to use data from other sources to inform a model on a target task. Current research has been focused on employing this methodology to real-world structures by using simulated structures for source information. This paper analyzes the feasibility of deploying this framework across multiple real-world structures. Data from two experimental scale models were evaluated in a multiclass damage detection problem. Damage in the structures was simulated through the removal of structural components. The dataset consists of the response from accelerometers equipped to the structures while the structures were under the influence of an external force. A convolution neural network (CNN) was used as the target-only model, and a Generative adversarial network (GAN) based CNN network was evaluated as the transfer learning model. The results show that transfer learning improves results in cases where limited data on the damaged target structure is available, however transfer learning is much less effective than traditional methods when there is a sufficient amount of data available.
Transfer Learning, Convolution Neural Network, Structural Health Monitoring.
Pekka Koskela, Anni Karinsalo, Jori Paananen and Laura Salmela, 1-4VTT Technical Research Centre of Finland Ltd, P.O. Box 1000, FI-02044 VTT, Finland
Since the mid-2000s, the digitalisation of border checks has often referred to the increased adoption of automated border control (ABC) solutions at border crossing points in all border environments from airports and seaports to land border crossings. Key prerequisites for the operational implementations of the so-called eGates have been the electronic machine-readable travel document together with biometric technologies that have facilitated the automation of much of the tasks performed by border guards at manual control booths for selected groups of nationalities. Now, the next wave of major changes is emerging with the development of electronic identification (eID), with certain implementations particularly designed for cross-border use cases supplementing and possibly replacing the traditional physical identity document in a long-term future. The evolution of eID strongly aligns with the increased demands for data privacy to ensure that individuals can better control how much information is shared about themselves, with whom and for what purpose. One possible technology to provide the so-called data self-sovereignty is distributed ledger technology(DLT), including blockchains. DLT is being developed for instance by the Linux foundation, dispensing several distributed ledger projects and associated solutions for digital and self-sovereign identity. One of these projects is Hyperledger Indy. In this study, we present a distributed ledger implementation based on Hyperledger Indy applied as a border check use case. Our aim is to investigate the suitability of DLT in providing data self-sovereign facility in border checks, and to discuss the benefits and disadvantages the technology might entail for this security domain.
Blockchain, Self-sovereign Identity, Border Control.
Ali Asad Sarfraz and Shiren Ye, Department of Computer sciences, Aliyun school of big data, changzhou University, china
This work details the byzantine fault-tolerant protocols that dynamically allow replicas to join and exit. Byzantine fault-tolerant (BFT) protocols and the blockchain now play an essential role in achieving consensus. There are numerous drawbacks to PBFT, despite its multiple positives. The first thing to note is that it runs in an environment completely isolated from the rest of the world. The entire system must be shut down before any nodes can be added or removed. Second, it ensures liveness and safety if no more than (n-1)/3 out of total n replicas, PBFT takes no action to cope with ineffective or malicious counterparts. This is bad for the system and will lead to its eventual failure. These flaws have far-reaching consequences in real life. The Randomization PBFT is an alternative way of dealing with these issues. In recent decades, as computer technology has advanced, so has our reliance on the products, services, and capabilities that computers provide.
DRBFT, PBFT, protocol, Blockchain, Byzantine fault tolerance, and Dynamic PBTF.
Tamara Al-Masri, Ahmed Al-Tamimi and Muawya Al-Dalaien, Department of Computer Science, Princess Sumaya University for Technology.
Agile Methodologies are one of the most effective methodologies, as it provides change and modification at each phase of the software life cycle according to the clients requirements. Still, it lacks security in those phases which makes it vulnerable to attacks or makes the project a liability to the housing corporate. The constantly changing customer demands, miscommunication, and misinterpretation between the task owner and the customer may lead to overwhelming the corporate’s resources and draining the project’s allocated time and budget. This can be mitigated by appending certain practices to the implemented development framework. In this paper, we candidate an enhanced Extreme Programming (XP) framework, one of the most common methodologies of Agile without the need to extend the process. These practices address the agonizing issue of redeveloping certain parts or all of the project and the neglect of the safety aspect throughout the development process.
Agile methodologies, Software development, Secure Software development, XP, Extreme programming.
Emmanuel Etuh1,2, Francis S. Bakpo1, George E. Okereke1 and Deborah U. Ebem1, 1Department of Computer Science, Faculty of Physical Sciences, University of Nigeria Nsukka, Nigeria, 2Department of Mathematics/Statistics/Computer Science, Kwararafa University, Wukari, Taraba State, Nigeria
Social media platform has greatly enhanced human activities in the virtual community. Virtual socialization has positively influenced social bonding among social media users irrespective of one’s location in the connected global village. Human user and social bot user are the two types of social media users. While human users personally operate their social media accounts, social bot users are developed software that manages a social media account for the human user called the botmaster. This botmaster in most cases are hackers with bad intention of attacking social media users through various attacking mode using social bots. The aim of this research work is to design an intelligent framework that will prevent attacks through social bots on social media network platforms.
Social media platform, human user, social bot, hackers, social security, intrusion prevention.
Zahraa Chreim1, Hussein Hazimeh1, Hassan Harb1, Fouad Hannoun2, Karl Daher2, Elena Mugellini2 and Omar Abou Khaled2, 1Lebanese University, Faculty of Science, Beirut, Lebanon, 2University of Applied Sciences of Western Switzerland, Fribourg, Switzerland
Search engines are among the most popular web services on the World Wide Web. They facilitate the process of finding information using a query-result mechanism. However, results returned by search engines contain many duplications. In this paper, we introduce a new content-type-based similarity computation method to address this problem. Our approach divides the webpage into different types of content, such as title, subtitles, body, etc. Then, we find for each type a suitable similarity measure. Next, we add the different calculated similarity scores to get the final similarity score between the two documents, using a weighted formula. Finally, we suggest a new graph-based algorithm to cluster search results according to their similarity. We empirically evaluated our results with the Agglomerative Clustering, and we achieved about 61% reduction in web pages, 0.2757 Silhouette coefficient, 0.1269 Davies Bouldin Score, and 85 Calinski Harabasz Score.
Information Retrieval, Websites Similarity, Graph Representation, Similarity Measures, Graph Kernel, Deduplication, Search Engines.
Indunil Ramadasa, Lahiru Liyanage, Theshan Dilanka and Dinesh Asanka, Department of Industrial Management, University of Kelaniya, Kelaniya, Sri Lanka
This research is an effort to analyze the effectiveness of Google Translations API (Google cloud) as a supportive tool for NLP practices as a solution for the lack of feasible Sinhalese language-specific NLP tools. Google Translation of Sinhalese to English and vice versa is being questioned for accuracy while significant improvement in the quality of translation has been observed over time. However, regardless of the perfection of translation output, it was observed that the same accuracy level might be sufficient for NLP techniques such as sentiment analysis and Named Entity Recognition (NER), where analysis is focused on individual words and/or/or sentences in most cases. It was possible to conclude that with our findings the use of Google Translation API as an intermediate supportive tool is giving reasonably well-acceptable aggregated results with Sentiments Analysis and a very high level of accuracy in terms of NER.
Natural Language Processing (NLP), Sinhalese, Sinhala, English, Sentiment Analysis, Named Entity Recognition (NER), Google Cloud Translation, Google Translation.
Hannah Cho, North Hollywood High School, 5231 Colfax Ave, North Hollywood, CA 91601
It is well known that meditation is an ef ective stress-relief technique. This can be attributed tomeditation’s ef ects on the autonomic nervous system; it activates the parasympathetic nervous system (PNS), and inhibits the sympathetic nervous system(SNS). Stress has been unprecedentedly exacerbated with the onset of the COVID-19 pandemic. This study investigates how meditation can be further personalized and adapted to enhance its ef ects on the autonomic nervous system, with a focus on teachers at a local progressive high school. Teachers wore physiological data-recording wristbands formonthly meditation sessions. Wristbands recorded electrodermal activity, blood volume pressure, heart rate, peripheral skin temperature, and motion-based activity. Collected wristband data was then exported into Kubios, a HRV analysis software, to calculate PNS and SNS indexes during the meditation portion and the interactive engagement that followed. Heart rate variability time-domain indicators that were used to calculate the indexes- Mean HR (bpm), Baevsky Stress Index, Mean RR (ms), RMSSD (ms)- and nonlinear Poincaré plot values were also compared. Focus meditations exerted higher PNS indexes and lower SNS indexes than during the engagement periods. Visualization meditations that called for active thinking exerted mixed and opposing results across participants, alluding that the type of vision participants are instructed to imagine may dif erently manipulate the ANS branches. This concept that meditation types can uniquely af ect nervous system control can be used to personalize meditation as clinical treatment. Although the ecological validity of this out-of-lab study is indeterminate, the findings of this study can lead to further research in laboratory settings.
Heart Rate Variability, Autonomic Nervous System, Stress, Meditation, Kubios.
Ali Rahama, Information Technology Specialist,6701 Seybold Rd STE 212, Madison, WI 53719, USA
The objective of this paper is to investigate the possible correlation between extinct Meroitic and local languages in todays Sudan. The researchers goal is implementing Natural Language Processing and use of phonetic comparison algorithms. Implementing phonetics to Natural language Processing is requiring great effort, our model produces both alphabetic mappings and translations of words into corresponding cognates implementing Soundex algorithm to capture character mappings and high-level morphemic correspondences. Researching this matter using Natural Language processing tools could lead to enhancement for Phonetic Comparison Algorithms in Low-Resources Languages area which currently lack researchers’ attentions.
Natural Language Processing (NLP), Phonetics, Meroitic, Phonemes, Phonetic, Soundex, Meroitic.
Shahajd and K. Mustafa, Department of Computer Science, Jamia Millia Islamia University, New Delhi, India
In the age of Apps, there has been a widespread proliferation of Learning Apps (LA). Almost every educational institution has been affected since the pandemic. The research indicates that such apps are highly effective for the so-called touch-screen generation in a variety of contexts. Data on LAs performance show that they are associated with compelling increases in student achievement. Recognizing the significance, it is suggested that teachers and other caretakers become involved in this new trend of mobile learning. Despite this, experts generally highlight issues concerning their effectiveness (see later). As a result, we observe the emergence of several design paradigms, having no or little theoretical bases. Even though businesses grow, and new tools and technology are developed, there isnt a good app design strategy based on accepted didactics. Realizing this, we suggest a Pedagogic-Object-Oriented-Based Approach to the design and development of LA, building on the idea of IEEE learning objects and the success of the object-oriented paradigm.
Learning App, App Ingredient, Object-based Approach, Design.
G. N. R. Prasad, Sr. Asst. Professor, Dept. of MCA, Chaitanya Bharathi Institute of Technology (Autonomous), Gandipet, Hyderabad – 500 075, India
A growing number of industries are using data mining to uncover knowledge that is concealed in vast data sets. However, there are also unfavourable societal conceptions regarding data mining, including the possibility of prejudice and privacy infringement. The latter involves treating someone unfairly due to their membership in a certain group. Making automated judgments, such as loan approval or rejection, insurance premium calculation, etc., is now possible thanks to automated data collecting and data mining techniques like categorization rule mining. Discriminatory choices may result if the training data sets are skewed in terms of traits like gender, colour, religion, etc. Data mining has been expanded to include antidiscrimination strategies, such as discrimination detection and prevention. Either direct or indirect discrimination is possible. Decisions based on delicate characteristics lead to direct discrimination. When judgments are formed based on non-sensitive traits that have a substantial correlation with biassed sensitive ones, indirect discrimination results. This article discusses how to prevent discrimination in data mining and offers fresh methods that may be used to prevent direct or indirect discrimination separately or together. Direct and indirect discriminatory decision rules are transformed into legal (nondiscriminatory) classification rules through the cleaning of the training data sets and outsourcing data sets. The suggested methods successfully eliminate both direct and indirect discriminatory biases from the original data set while maintaining the integrity of the data.
Antidiscrimination, data mining, direct and indirect discrimination prevention, rule protection, rule generalization, Privacy.
Bibekananda Panda and Poonam Singh, Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, India
The non-orthogonal multiple access (NOMA) approaches have recently received much attention. It has also been a potential method for wireless communication systems beyond the fifth generation (5G). The successive interference cancellation (SIC) procedure in NOMA systems is often carried out at the receiver, where several users are sequentially decoded. The successful detection of prior users will significantly influence the detection accuracy due to the effects of interferences. A deep learning-based NOMA receiver is analyzed to detect signals for multiple users in a single application without determining channels. This paper analyzes deep learning (DL)- based receiver for NOMA signal detection concerning several DL-aided sequence layers and optimizers by training orthogonal frequency division multiplexing (OFDM) symbols. The simulation outcomes illustrate the various DL-based receiver characteristics using the traditional SIC approach. It also demonstrates that the accuracy of the different DL-based models is more predictable than the SIC approach.
NOMA, DNN, GRU, LSTM, Bi-LSTM.
Arunima Borah, Faculty of Engineering and Technology, MSRUAS, Bengaluru, India
Creating fake videos has become easy due to the advancement ofdeepfake technology. Upon closer analysis, severalaudio and video modifying algorithms disclose artefacts, that is parallel to conventional Computer Vision use-cases that emanates from face tracking and voice recognition architypes. The sole attempt of this study is to focus on the uprising interest of doctored media across the global population and how Neuro-Science can act as a blessing when amalgamated with AI while concluding with an efficient research direction in future involvement of ‘Explainable-AI’ in computer vision.
Deepfakes, Neuro-Science, Visual Cocktail Party Effect, Long-Short Term Memory, ResNext101.
Goutam Kumar Sahoo, Rashmiranjan Nayak, K. L. Sanjeev Tudu, Umesh Chandra Pati, Santos Kumar Das and Poonam Singh, Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Odisha, India
Road accidents are widespread in this fast-track lifestyle where time is precisely equal to money. Road accidents are prone at sharp corners, where roads behind the turning lie in the blind spot. Research says 30% of road accidents only occur due to blind spot problems. This work aims to detect moving vehicles using unsupervised methods in videos and estimate their speed and distance using cameras mounted in road infrastructure for collision avoidance at sharp corners. This Internet of Things-enabled edge computing system development and installation near the blind spot in a safe space will help ensure the safety and security of road users without any in-vehicle sensors. Information like the availability of vehicles in the area lying on the blind zone, speed, and distance of the upcoming vehicle can be shared with the drivers beforehand for safety purposes. An edge computing platform-based visual indication and an alarm generation will guide the driver by a fixed roadside unit (RSU) near the turning point. The vehicle detection method uses a lightweight computer vision algorithm comprised of four stages to find the coordinates of moving objects in two successive frames, starting with frame differencing, Image thresholding, Image dilation, and finding contours. An image-based vehicle speed estimation is performed using the concept of pixels per meter (PPM). The proposed technique is verified and tested using a laboratory setup on different videos. Thus, the performance of the system enables the vehicle to prevent a collision. The importance of vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) communication will strengthen the safety of road users.
Blind Corners, Camera, Edge computing, Internet of Things, Vehicle Detection, Warning Generation, V2I (Vehicle-to-infrastructure) Communication.
Vidyalekshmi Chandrikal and Simi Surendran, School of Engineering, Amritapuri, Kollam, Kerala, India
This paper gives complete guidelines for authors submitting papers for the AIRCC Journals. One of the significant health problems around the globe is associated with neonatal mortality or morbidity and disability in later life. The primary cause of neonatal death is preterm labor, which accounts for less than a half percent of all deaths among children under five years. Continuous monitoring and risk prediction could help provide medical assistance to the pregnant woman at the right time, substantially reducing neonatal mortality. We conducted a detailed data analysis and comparative study on various machine learning models on the cardiotocography dataset to conclude on a better accuracy of preterm birth prediction. This paper proposes an incremental learning approach to predict preterm labor risk. The response time for the medical aid can be significantly reduced using this incremental edge learning.
Preterm Birth, Machine Learning, Fetal Health Monitoring, Incremental Learning.
Esraa Fady Dawood1, Nader Mahmoud2 and Ashraf Elsisi3, 1Department of Computer Science, Faculty of Computers and Information, Menoufia University, Egypt, 2Department of Computer Science, Faculty of Computers and Information, Menoufia University, Egypt, 3Department of Computer Science, Faculty of Computers and Information, Menoufia University, Egypt
COVID-19 caused a pandemic with countless deaths. From the outset, clinical institutes have explored chest Computed Tomography (CT) as an effective and complementary screen. On CT images, artificial intelligence approaches are utilized to accurately detect diseases, and this work intends to design a methodology that can diagnose COVID-19 using deep learning methods on CT images. To categorize patients infected and non-infected with COVID-19, we utilized the Mask R-CNN approach to train and test on the dataset. In this study, the data classes were restructured using the Fuzzy Color technique as a preprocessing step and the images that were structured with the original images were stacked. The dataset presented here provides a significant number of frontal CT image views, which are an important resource for the algorithms employed in the development of COVID-19 detection tools. Using chest CT images, the proposed model has a precision of 97.88 % and an accuracy of 98.28 %. The entire procedure is described in great detail. When a comparison table of AI-based techniques is created, it becomes clear that the Mask R-CNN methodology on chest CT images performs better in terms of COVID-19 identification. COVID-19 detection from chest CT images is proven to be accurate and resilient using the Mask R-CNN approach.
COVID-19, chest radiography, Fuzzy color technique, mask R-CNN, Computed Tomography (CT).
Rey Aliño1 and Proceso Fernandez2, 1Ateneo de Davao University, Davao City, Philippines, 2Ateneo de Manila University, Quezon City, Philippines
The proliferation of Invasive Alien Species (IAS) in the Philippines is a major threat to its biodiversity. Taxonomic information on IAS is crucial to preventing, controlling and eradicating it. Computer vision technology can be applied to assist in strategies and plans to fight IAS. Invasive alien species in the Philippines (IAS PH) include five alien frogs: the cane toad (Rhinella marina), Chinese bullfrog (Hoplobatrachus rugulosus), green paddy frog (Hylarana erythraea), greenhouse frog (Eleutherodactylus planirostris), and Asiatic painted toad (Kaloula pulchra). This study presents an implementation of ResNet50, a well-known convolutional neural network (CNN) model, previously used for other computer vision tasks, for classifying the five invasive alien frog species in the Philippines and distinguishing them from endemic Philippine frogs. In this interdisciplinary study, a dataset of 252 images of the five alien frogs collected by experts from the ASEAN Centre for Biodiversity was used. A sixth endemic class was added containing 304 endemic frog images gathered from the International Union for Conservation of Nature (IUCN) database, the Global Biodiversity Information Facility (GBIF), ACB and the web. The images were pre-processed and used to re-train ResNet50 to classify the five alien frog classes and the endemic frog class. We used five-fold cross validation to evaluate the performance of our model. Precision, recall, F1-score and overall model accuracy metrics show that the ResNet50 CNN model can fairly classify invasive and endemic frog classes from each other. An overall network accuracy rate of 92% was achieved by the ResNet50 model. Increasing the currently small dataset of frogs endemic and invasive to the Philippines, though, is recommended to improve the accuracy of the model.
Invasive Alien Frog Species, Endemic Philippine Frog Species, Convolutional Neural Networks, Biodiversity.
Ashok Yadav and Vrijendra Singh[0000−0002−8818−5673], Department of Information Technology, Indian Institute of Information Technology, Allahabad, India
Publicly available information on social networks helps in the analysis of crises, natural disasters, etc. Social networks are an important medium that help in sharing information during various crises. Russia- Ukraine is currently one of the ongoing crises. The data we collect during any crisis helps us in analyzing public opinion and their sentiments. In this work, we have collected Twitter data using the Twint module to understand the sentiments shared by the public during the Russia-Ukraine crisis. For sentiment analysis, we have used various machine learning classifiers like support vector machine, Random Forest, k-nearest neighbors (KNN) and (Xgboost). The training and testing accuracy of the model with the Xgboost classifier is 97% and 99% respectively and the loss is 0.18.
SVM · KNN · Random Forest · Sentiment Analysis · Twitter Data.
Zeinab Khanjarinezhadjooneghani, Joseph Wireman and Nasseh Tabrizi, Department of Computer Science, Greenville, North Carolina
The COVID-19 pandemic is one of the greatest challenges we are facing as a global society. This study aims to determine concerns towards the COVID-19 vaccine using Twitter data. This study will be comparing four topic modelling methods, the Latent Dirichlet Allocation (LDA), LDA with Gibbs Sampling (LDA-MALLET), and Non-Negative Matrix Factorization (NMF) as well as Top2Vec methods. The data comes from a series of tweets from December 16th, 2020, to January 10th, 2021. These models will be compared based on Coherence score as well as Jaccard score. The LDA-MALLET method was best in terms of coherence score and was followed by NMF and LDA respectively. In terms of Jaccard score, LDAs was superior to NMF. Top2Vec returned 252 topics, after merging similar topics, the average Coherence score of topics was better than the LDA and NMF method and the average Jaccard score was better than the NMF model.
Social Media Analysis, Twitter Data, COVID-19 Vaccine, Topic Modelling, LDA, LDA-MALLET, NMF, Top2Vec, Data Analysis, Natural Language Processing
Rachid Sabre, Laboratory Biogéosciences CNRS, University of Burgundy/Agrosup Dijon, France
In this work, we consider a symmetric harmonizable stable random field (two-dimensional signal) with certain mixture spectral density observed with an additive constant error. This paper gives an estimator of the constant error and its rate of convergence. We study the rate of convergence when the spectral density have some behaviours at origin. Few long memory signals are taken here as example.
Spectral density, Jackson kernel, Stable random fields.
Prajit Nadkarni and Narendra Varma Dasararaju, Flipkart Internet Pvt. Ltd
Visual search is of great assistance in reseller commerce, especially for nontech savvy users with affinity towards regional languages. Product attributes available in e commerce have potential for building better visual search systems [2, 20, 29]. We design a visual search system for reseller commerce using a multitask learning approach and address challenges like image compression, cropping, etc, faced in reseller commerce. Our model consists of three tasks: attr ibute classification, triplet ranking and variational autoencoder (VAE). We introduce an offline triplet mining technique which utilizes information from multiple attributes to capture relative order within data. This technique displays better performance compared to traditional triplet mining  baseline. We compare and report incremental gain achieved by our unified multi task model over each individual task separately. The effectiveness of our method is demonstrated using in house dataset of images fro m the Lifestyle business unit of Flipkart. To efficiently retrieve images in production, we use Approximate Nearest Neighbor (ANN) index.
Content based image retrieval, Visual search, Multi task Learning, Triplet loss, Variational autoencoder.
Esmael Hamuda1, Ashkan Parsi2, Martin Glavin2 and Edward Jones2, 1Department of Electrical and Computer Engineering, Elmergib University, Al Khums, Libya, 2Department of Electrical and Electronic Engineering, University of Galway, Galway Ireland
In this paper, we investigate the use of deep learning approaches for plant classification (cauliflower and weeds) in smart agriculture applications. To perform this, five approaches were considered, two based on well-known deep learning architectures (AlexNet and GoogleNet), and three based on Support Vector Machine (SVM) classifiers with different feature sets (Bag of Words in L*a*b colour space, Bag of Words in HSV colour space, Bag of Words of Speeded-up Robust Features (SURF)). Two types of datasets were used in this study: one without Data Augmentation and the second one with Data Augmentation. Each algorithms performance was tested with one data set similar to the training data, and a second data set acquired under challenging conditions such as various weather conditions, heavy weeds, and several weed species that have a similarity of colour and shape to the crops. Results show that the best overall performance was achieved by DL-based approaches.
Deep Learning, BoWs, SURF, Data Augmentation, Plant Classification and Smart Agriculture.
Mohammad Saleem and BenceKovari, Department of Automation and Applied Informatics, Budapest University of Technology and Economics, Budapest, Hungary
Online signatures are one of the most commonly used biometrics. Several verification systems and public databases were presented in this field. This paper presents a combination of k-nearest neighbor and dynamic time warping algorithms as a verification system using the recently published DeepSignDB database. Our algorithm was applied on both finger and stylus input signatures which represent both office and mobile scenarios. The system was first tested on the development set of the database. It achieved an error rate of 6.04% for the stylus input signatures, 5.20% for the finger input signatures, and 6.00% for a combination of both types. The system was also applied to the evaluation set of the database and achieved very promising results, especially for finger input signatures.
Online signature verification, k-nearest neighbor, dynamic time warping.
Sayed Ahmad Salehi, Department of Electrical and Computer Engineering, University of Kentucky, Lexington, Kentucky, USA
Stochastic computing (SC) provides a fault-tolerant and low-cost alternative to conventional binary computing (BC). The capacity of SC to implement complex mathematical functions with simple logic gates creates a path toward the design of efficient hardware architectures. This paper presents a new methodology for the hardware implementation of servomotor controller using SC. We design SC circuits using both quadrature decoder and efficient decoder for implementing servo controller and compare them with traditional BC-based servo controller. The quadrature decoder requires more hardware resources than efficient decoder but can provide position information in PWM form. The FPGA implementation result shows that, compared to BC-based design, quadrature decoder-based design achieves 56.7% savings in area and 33.33% savings in power consumption, and efficient decoder-based design achieves 73.7% savings in area and 33.33% savings in power consumption.
Stochastic computing (SC), servomotor controller, quadrature decoder, efficient decoder, stochastic integrator.
Zainab Albujasim1, Diana Inkpen2 and Yuhong Guo1, 1School of Computer Science, Carleton University, Ottawa, ON, Canada, 2School of Electrical Eng. and Computer Science, University of Ottawa, ON, Canada
Word embedding is the foundation of modern language processing (NLP). In the last few decades, word representation has evolved remarkably resulting in an impressive performance in NLP downstream applications. Yet, word embedding’s interpertablity remains a challenge. In this paper, We propose a simple technique to interpret word embedding. Our method is based on post processing technique to improve the quality of word embedding and reveal the hidden structure in these embeddings. We deploy Co-clustering method to reveal the hidden structure of word embedding and detect sub-matrices between word meaning and specific dimensions. Empirical evaluation on several benchmarks shows that our method achieves competitive results compared to original word embedding.
Word Embedding, Quantization, Post-processing.