Search

Academic studies on the Automatic Voltage Regulator (AVR) have focused on a linear mathematical model which lacks protective features. A more accurate model of the AVR system includes protective features as described in IEEE standards. The AVR models without protective features, namely limiters, result in less accurate control performance since the outputs of real controllers are always bounded. In the present study, the controller outputs are limited between -0.9pu and 1.0pu, and the upper bound of exciter output is limited by 3.1pu with the direction of IEEE standards. The effect of limiters on the controller performance is investigated. Two controllers with a novel sliding surface function are proposed based on a mathematical model of the AVR and its approximate reduced-order model. The proposed controllers having only two parameters are optimized with improved particle swarm optimization (PSO) algorithm. After optimization, the robustness of the controllers is compared with the results of various operating conditions identical to the previous studies. Although controller inputs and exciter outputs are limited, the maximum overshoot is measured as less than 0.1% at no-load conditions. Nominal time constants of the AVR constituents are perturbed from -50% to 50%, and ±10% load disturbance is applied to the output. The robustness of the controllers against parameter uncertainties is measured with an average overshoot at the output. Another superiority among the reported results in the literature is obtained from the proposed controllers, where the minimum average overshoot rates are obtained. In addition, when ±10% load is applied to the output of the AVR, the proposed controllers generate accurate control inputs to reject the load disturbance successfully. Furthermore, the proposed controllers keep the output within the ±5% band if there is a monotonic change at the output. All the results show that the proposed controllers with the improved PSO have drawn the best performance from the perspective of time-domain specifications in comparison with the recently reported controllers.

Due to the development of technology and organizational needs, the competence of the State Civil Apparatus has increased significantly, necessitating the adjustment of improvement techniques. Furthermore, the lack of a role for the sending agency in training continues to be an impediment to the intersperity between competence needs and the implementation system. This research focuses on how the competency-based training model is applied, as well as the function and impact of supervisory authorities' training results on enhancing performance. This study uses a combination of qualitative and quantitative methods to obtain results that are relevant to the needs, and it was conducted at the State Administration Agency's Center for Training and Development and Competency Mapping of State Civil Apparatus with alumni research training supervisory officials in 2018, 2019, and 2020. According to the findings of the study, the training has not taken advantage of technological advancements or maximized the role of the sending agency, so there are still difficulties and limits in its execution, particularly in terms of location and time. To meet these issues, a training model must be developed that focuses on ensuring the organizer's competency.

Currently, several universities are using the learning management system (LMS) application for administration, documentation, activity reporting, learning activity implementation, and online assignments. SIAKAD is a learning management system application that has been implemented at STIKOM Poltek Cirebon. At SIAKAD, students and lecturers can download and upload materials, view course data and schedule data. However, in SIAKAD there are several features that are still not fulfilled, such as checking attendance data and recapitulation, task recapitulation and automation of calculating final course grades. In this study the Lemskom Poltek Cirebon application was designed using the Asynchronous Learning method. Which aims to complement features that do not exist in SIAKAD. In the Lemskom Poltek Cirebon application, which is designed using the Asynchronous method, it is very appropriate because this method is not based on time and place and is very good at helping in the learning process. So that students can see for themselves the course schedule, their attendance data, and the assignments given by the lecturer at anytime and anywhere along with the information and values they get. Lecturers can see the recapitulation of attendance, and recapitulation of assignments as well as details of what grades are in the courses they teach in more detail and can calculate the final grade automatically.

The bilingual experience may place special cognitive demands on speakers and has been argued to lead to improvements in domain-general executive abilities, like cognitive control and working memory. Such improvements have been argued for based on both behavioral and brain imaging evidence. However, the empirical landscape is complex and ridden with controversy. Here we attempt to shed light on this question through an fMRI investigation of relatively large, relatively homogeneous, and carefully matched samples of early balanced bilinguals (n = 55) and monolinguals (n = 54), using robust, previously validated individual-level markers of neural activity in the domain-general multiple demand (MD) network, which supports executive functions. We find that the bilinguals, compared to the monolinguals, show significantly stronger neural responses to an executive (spatial working memory) task, and a larger difference between a harder and an easier condition of the task, across the MD network. These stronger neural responses are accompanied by better behavioral performance on the working memory task. We further show that the bilingual-vs.-monolingual difference in neural responses is not ubiquitous across the brain as no group difference in magnitude is observed in primary visual areas, which also respond to the task. Although the neural group difference in the MD network appears robust, it remains difficult to causally link it to bilingual experience specifically.

International audience

Pollen allergies are a growing concern for human health. This is why automated pollen monitoring is becoming an important area of research. Machine learning approaches show great promise for tackling this issue but these algorithms need large training data sets to perform well. This study introduces a new pollen data set, obtained using a Rapid-E particle analyzer, that is representative for the flora of Romania. Pollen, from thirteen species present in Romania, was used in developing this database with over 100 thousand samples measured. Our study shows performance similar to or above that of humans in the task of pollen classification on the newly introduced data set using a convolutional neural network.

Automatic speech recognition (ASR) for Romanian language is on an ascending trend of interest for the scientific community. In the last two years several research groups reported valuable results on speech recognition and dialogue tasks for Romanian. In our paper we present the improvements we recently obtained by collecting and using more text and audio data for training the language and acoustic models. We emphasize the automatic methodologies employed to facilitate data collection and annotation. In comparison to our previous work, we report state-of-the-art results for read speech (WER of 1.6%) and significantly better results on spontaneous speech: relative improvement around 40%). In order to facilitate direct comparison with other ASR systems, we release all evaluation datasets, totaling 10 hours of manually annotated speech.

International audience; The increased complexity of programs and pro-cessors is an important challenge that the embedded real-time systems community faces today, as it implies substancial timing variability. Processor features like pipelines or communication buses are not always completely described, while black-box programs integrated by third parties are hidden for IP reasons. This situation explains the use of statistical approaches to study the timing variability of programs. Most existing work is concentrated on the guarantees provided by positive answers to statistical tests, while our current work concerns potential algorithms based on the negative answers to these tests and their impact on the timing analysis. We introduce here one such algorithm, the Walking Kolmogorov-Smirnov test (WKS).

International audience; Designers of embedded real-time systems derive, in general, their time parameters such as activation periods from those of sensors or actuators. By designers, we mean the team in charge of conceiving embedded real-time systems. This team includes Control Theory designers and Computer Science designers. Within this paper we present the point of view of Computer Science designers, while the periods proposed by Control Theory designers are supposed robust with respect to the physical behavior of the system. The execution times are, then, estimated by studying statically the programs structure or dynamically the programs execution. In some cases, both activation periods and execution times depend on a sensor information. For instance, they depend on the angular speed of wheels within an automotive embedded real-time system and such systems follow a rate-dependent model. Elastic tasks is another model, where one may consider execution time variation depending on the selected period. Within this paper, we are interested in describing statistically the relationship between activation periods and execution times of programs. More precisely, we study the impact of the period variation on the distributions of the execution times. To illustrate our preliminary results, we consider, as case study, the set of programs executing the autopilot of an open-source PX4 drone.

To fully unleash the potential of graphene-based devices for neuromorphic computing, we propose a graphene synapse and a graphene neuron that form together a basic Spiking Neural Network (SNN) unit, which can potentially be utilized to implement complex SNNs. Specifically, the proposed synapse enables two fundamental synaptic functionalities, i.e., Spike-Timing-Dependent Plasticity (STDP) and Long-Term Plasticity, and both Long-Term Potentiation (LTP) and Long-Term Depression (LTD) can be emulated with the same structure by properly adjusting its bias. The proposed neuron captures the essential Leaky Integrate and Fire spiking neuron behavior with post firing refractory interval. We demonstrate the proper operation of the graphene SNN unit by relying on a mixed simulation approach that embeds the high accuracy of atomistic level simulation of graphene structures conductance within the SPICE framework. Subsequently, we analyze the way graphene synaptic plasticity affects the behavior of a 2-layer SNN example consisting of 6 neurons and demonstrate that LTP significantly increases the number of firing events while LTD is diminishing them, as expected. To assess the plausibility of the graphene SNN reaction to input stimuli we simulate its behavior by means of both SPICE and NEST, a well established SNN simulation framework, and demonstrate that the obtained reactions, characterized in terms of total number of firing events and mean Inter-Spike Interval (ISI) length, are in close agreement, which clearly suggests that the proposed design exhibits a proper behavior. Further, we prove the unsupervised learning capabilities of the proposed design by considering a 2-layer SNN consisting of 30 neurons meant to recognize the characters “A,” “E,” “I,” “O,” and “U,” represented with a 5 by 5 black and white pixel matrix. The SPICE simulation results indicate that the graphene SNN is able to perform unsupervised character recognition associated learning and that its recognition ability is robust to input character variations. Finally, we note that our proposal results in a small real-estate footprint (max. 30 nm$^2$ are required by one graphene-based device) and operates at 200 mV supply voltage, which suggest its suitability for the design of large-scale energy-efficient computing systems.

The paper proposes a methodology for estimation of the variation of power devices lifetime using data from different stages of development and tests. In the characterization process of power devices, the estimation of the minimum lifetime is not a simple task, especially when each stress-test lasts weeks and only a few lifetime measurements are available. In this case, the estimation of the minimum lifetime based only on the distribution of measured data is not effective, so very large safety margins are taken for guaranteeing of the lifetime parameters. The proposed methodology overcomes the problem of limited available data. It determines, in an automatic way, the most relevant electrical parameters (measured before the stress-test) which influence the lifetime spread and then, uses their distributions from back-end stage, where thousands of devices are measured, to predict the distribution of the lifetime and, out of it, the minimum lifetime. Thus, by using the initial electrical parameters in the estimation of the lifetime spread, the variation of the manufacturing process is indirectly taken into account. The validation of the lifetime variation model, with leave-one-out method, shows a maximum relative error of 25%. With the proposed methodology, based on data from different stages of development, the minimum lifetime of power devices can be more accurately predicted, even when only a few lifetime measurements are available.

Data is a very authentic evidence that need to be saved because the data are at risk of data changes by the party who wanted to misuse the data or no effect on the data. Therefore, so that no unauthorized person to change the existing data, stored or can take the data that is very important, we need a method to be able to secure and keep data. The database is used to keep data and information. The whole system stores the data in the database, so that the contents of stored data security and confidentiality must be maintained. To maintain the security of a security that is required cryptography. With the use of databases that are in can be connected to each other, Database is not just a place to store data,Computer security system is needed at this time in line with the increase use of computers in the world. So existing data security has become seriously threatened to be accessed by people who are not responsible. Computer security becomes very important at this time because it is related to the privacy, integrity, authentication and confidentiality. Thus data need to be stored by using the correct storage techniques, so that the confidentiality and security of data is maintained. Cryptographic techniques with engineering data encryption on the database is one solution that can be used to meet the confidentiality and Data Security settings. With the existence of a cryptography that includes the encryption process, the data in the database can be encoded so that unauthorized persons can not read the information, in addition to those who know the key to describing it. Security of data / information document with a cryptographic technique that uses the Blowfish algorithm chosen method for cryptographic key modernmerupakan shaped symmetric block cipher. Blowfish algorithm which can encrypt text built in text form. Encryption is done using a specific key, resulting ciphertext that can not be understood. The ciphertext can be changed back as the beginning if the decryption using the same key in the beginning when the database mengekripsi.

A natural evolution of applications that analyze speech is to improve their robustness to multi-speaker environments. Humans use selective auditory attention and can easily switch focus from one source to another even when listening to a single channel recording with overlapped speech. The same brain feature allows us to detect the number of simultaneously active sources. In order to quantify human level performance for this task we have designed a perception study that evaluates participants’ ability to accurately count multiple speakers in a single channel audio file. We also analyzed the influence of listening time and of hearing familiar voices. The study was carried in 3 sessions, with the help of 31, 38, and 80 volunteers and significantly extends the findings in existing literature. Using the conclusions from the perception analysis, several convolutional neural networks were trained to estimate the number of competing speakers on speech timeframes ranging from 25 ms to 1 s. The same models were instructed to tag overlapped speech and we observed F-score values up to 0.91. For both tasks, the proposed methods lead to lower error than existing approaches and require smaller timeframes. Compared with human listeners, the neural networks can count speakers more accurately by analyzing a considerably shorter recording.

The data processing of the election of citizens entitled to get the public health card aid fund in Kaliwadas Subdistrict, Sumber Subdistrict of Cirebon Regency generally still use manual selection system, that is, there is no computerization in determining the people who are eligible to get the help of public health insurance card so that many problems happened at this system. The problem that often arises because the target of the recipients of the public health card is not right on target to the eligible citizens, such as people who are actually not eligible to get grants but get the grant, otherwise the less fortunate people who are entitled to receive public health card security card but did not get donation. From these problems need a system for decision making, a system that can provide recommendations as a consideration for decision making accurately and accurately. This system can support decision-making of candidates receiving health insurance cards based on predetermined criteria. In order to calculate the decision support system more accurately then used a method, namely method of Fuzzy Technique for Order Preference by Similiary to Ideal Solution (Fuzzy TOPSIS) to determine the citizens who are entitled to get public health insurance card and get the best alternative value.

Pollen allergies are a cause of much suffering in an increasing number of individuals. Current pollen monitoring techniques are lacking due to their reliance on manual counting of pollen by human technicians. In this study, we present a neural network architecture capable of distinguishing pollen species using data from an automated particle measurement device. This study presents an improvement over the current state-of-the-art in the task of automated pollen classification using the fluorescence spectrum of aerosol particles.

The last decade brought significant advances in automatic speech recognition (ASR) thanks to the evolution of deep learning methods. ASR systems evolved from pipeline-based systems, that modeled hand-crafted speech features with probabilistic frameworks and generated phone posteriors, to end-to-end (E2E) systems, that translate the raw waveform directly into words using one deep neural network (DNN). The transcription accuracy greatly increased, leading to ASR technology being integrated into many commercial applications. However, few of the existing ASR technologies are suitable for integration in embedded applications, due to their hard constrains related to computing power and memory usage. This overview paper serves as a guided tour through the recent literature on speech recognition and compares the most popular ASR implementations. The comparison emphasizes the trade-off between ASR performance and hardware requirements, to further serve decision makers in choosing the system which fits best their embedded application. To the best of our knowledge, this is the first study to provide this kind of trade-off analysis for state-of-the-art ASR systems.

This paper consists in developing and testing new methods for historic building rehabilitation regarding thermal and light comfort for its occupants to minimize the building carbon footprint. A study was carried out on the degree of comfort in a historic building that requires innovative solutions for controlling indoor temperatures, minimally invasive on the outside architecture.

The present study proposes the product design of an innovative external luggage storage system, analyzed from economical point of view which can be digitalized by developing the attached module in order to organize the distribution of the passengers and luggage in the same time. It has been developed a case study of an industrial product of which assembly uses standardized structural materials among side modular subassemblies used in railway industry. The predicted cost effectiveness of a product is an essential part of the design process that will influence the entire workflow of the product design cycle. It consists in detailed structure analysis of each part, manufacturing process and assembly methods from a cost point of view. The product studied is a modular structure that can provide secure user-friendly storage space in public spaces. The cost analysis will provide a full insight in the feasibility of the product from the customer point of view. The product contains multiple modules in order to respect the ergonomic requirements and to increase the comfort of the passengers. The main advantage of this innovative baggage system is that it was designed to be fixed on the external structure of the wagon. Thus, the passenger concerns about the safety of the luggage ends at the beginning of the travel trip; the insufficient storage space designed in interior of the wagon is assured outside, it also solves the problem of difficult access for the elderly and kids accessing the storage area inside the wagon, especially when the train is moving and decreases the time required for embarking and disembarking for passenger with very big luggage. Thus, such a product would fluidize the travel time flow of the train.

The paper presents aspects related to the operation of an electronically controlled diesel engine that equips two cars. Some data related to 100 experimental tests for each car are presented. Some elements related to the electronic control and stability of the Diesel engine are presented by using the fractional calculation to generalize the study.

The SincNet architecture [1] was recently introduced as an approach to the speaker recognition task. Its main innovation was the sinc layer—an elegant and lightweight way of extracting features from speech. Despite good performance on multiple datasets, little information was provided on the architectural choices. In this work, we aim to shed some light on the importance of the network topology and various hyperparameters. We replace the original network trunk with a lightweight trunk inspired from residual networks (ResNets) and optimize its hyperparameters. Furthermore, we carry an extensive study on the sinc layer’s hyperparameters. Our main finding is that the stride and window size of the feature extractor plays a crucial role in obtaining good performance. Further experiments on conventional features, such as MFCCs and FBANKs, yield similar conclusions; in fact, by using optimal values for these two hyperparameters, traditional features are able to match the performance of sinc features. Surprisingly, the best results obtained go against conventional wisdom: an analysis window of only a couple of milliseconds and a stride of only a couple of samples are found to give the best results. Our code is available at https://bitbucket.org/doneata/sincnet.

Quantifying the confidence (or conversely the uncertainty) of a prediction is a highly desirable trait of an automatic system, as it improves the robustness and usefulness in downstream tasks. In this paper we investigate confidence estimation for end-to-end automatic speech recognition (ASR). Previous work has addressed confidence measures for lattice-based ASR, while current machine learning research mostly focuses on confidence measures for unstructured deep learning. However, as the ASR systems are increasingly being built upon deep end-to-end methods, there is little work that tries to develop confidence measures in this context. We fill this gap by providing an extensive benchmark of popular confidence methods on four well-known speech datasets. There are two challenges we overcome in adapting existing methods: working on structured data (sequences) and obtaining confidences at a coarser level than the predictions (words instead of tokens). Our results suggest that a strong baseline can be obtained by scaling the logits by a learnt temperature, followed by estimating the confidence as the negative entropy of the predictive distribution and, finally, sum pooling to aggregate at word level., Comment: Accepted at SLT 2021

In the paper we propose a reconfigurable graphene-based Spiking Neural Network (SNN) architecture and a training methodology for initial synaptic weight values determination. The proposed graphene-based platform is flexible, comprising a programmable synaptic array which can be configured for different initial synaptic weights and plasticity functionalities and a spiking neuronal array, onto which various neural network structures can be mapped according to the application requirements and constraints. To demonstrate the validity of the synaptic weights training methodology and the suitability of the proposed SNN architecture for practical utilization, we consider character recognition and edge detection applications. In each case, the graphene-based platform is configured as per the application tailored SNN topology and initial state and SPICE simulated to evaluate its reaction to the applied input stimuli. For the first application, a 2-layer SNN is used to perform character recognition for 5 vowels. Our simulation indicates that the graphene-based SNN can achieve comparable recognition accuracy with the one delivered by a functionally equivalent Artificial Neural Network. Further, we reconfigure the architecture for a 3-layer SNN to perform edge detection on 2 grayscale images. SPICE simulation results indicate that the edge extraction results are close agreement with the one produced by classical edge detection operators. Our results suggest the feasibility and flexibility of the proposed approach for various application purposes. Moreover, the utilized graphene-based synapses and neurons operate at low supply voltage, consume low energy per spike, and exhibit small footprints, which are desired properties for largescale energy-efficient implementations.

Design and implementation of artificial neuromorphic systems able to provide brain akin computation and/or bio-compatible interfacing ability are crucial for understanding the human brain’s complex functionality and unleashing brain-inspired computation’s full potential. To this end, the realization of energy-efficient, low-area, and bio-compatible artificial synapses, which sustain the signal transmission between neurons, is of particular interest for any large-scale neuromorphic system. Graphene is a prime candidate material with excellent electronic properties, atomic dimensions, and low-energy envelope perspectives, which was already proven effective for logic gates implementations. Furthermore, distinct from any other materials used in current artificial synapse implementations, graphene is biocompatible, which offers perspectives for neural interfaces. In view of this, we investigate the feasibility of graphene-based synapses to emulate various synaptic plasticity behaviors and look into their potential area and energy consumption for large-scale implementations. In this article, we propose a generic graphene-based synapse structure, which can emulate the fundamental synaptic functionalities, i.e., Spike-Timing-Dependent Plasticity (STDP) and Long-Term Plasticity . Additionally, the graphene synapse is programable by means of back-gate bias voltage and can exhibit both excitatory or inhibitory behavior. We investigate its capability to obtain different potentiation/depression time scale for STDP with identical synaptic weight change amplitude when the input spike duration varies. Our simulation results, for various synaptic plasticities, indicate that a maximum 30% synaptic weight change and potentiation/depression time scale range from [-1.5 ms, 1.1 ms to [-32.2 ms, 24.1 ms] are achievable. We further explore the effect of our proposal at the Spiking Neural Network (SNN) level by performing NEST-based simulations of a small SNN implemented with 5 leaky-integrate-and-fire neurons connected via graphene-based synapses. Our experiments indicate that the number of SNN firing events exhibits a strong connection with the synaptic plasticity type, and monotonously varies with respect to the input spike frequency. Moreover, for graphene-based Hebbian STDP and spike duration of 20ms we obtain an SNN behavior relatively similar with the one provided by the same SNN with biological STDP. The proposed graphene-based synapse requires a small area (max. 30 nm 2 ), operates at low voltage (200 mV), and can emulate various plasticity types, which makes it an outstanding candidate for implementing large-scale brain-inspired computation systems.

International audience; Real-time embedded systems (automotive, avionics,drones autopilots, etc.) are composed of numerous functional components that are continuously interacting via a variety of communication models where they intensively share data. For the overall functional correctness these systems must verify not only real-time scheduling requirements but they also must guarantee that the data being used are qualitatively correct. The quality of the data reflects the preservation of data related properties: temporal properties (i.e freshness, end-to-end latency, etc.) and integrity related properties (i.e data consistency). The proposed protocols to ensure such properties highly depend on the considered communication model (shared registers or large buffers) and the data access policy (directly or via local copies). In order to overcome this limitation, we provide in this paper, the means for managing the FIFO buffers to guarantee these data properties in a way that communication dependencies do not impact the tasks system scheduling order. We do so while considering the communication model presented in [14]. We provide methods computing a sufficient size for the considered buffers. Last but not least, an algorithm implementing the "last reader tags mechanism" together with the data temporal matching is provided and explained.

Measurement-based Worst Case Execution Time (WCET) estimations require appropriate benchmarks that are composed both by programs and representative measurement protocols. The representativeness is defined with respect to both the programs and processors executing them. In this paper we consider the representativeness with respect to a multicore processor and we provide preliminary conclusions on how appropriate benchmarks could be proposed.

The increased use of statistical libraries within the real-time community is facing today the lack of appropriate libraries. Many of them exist in R, but existing Python statistical libraries are lacking, particularily with respect to Extreme Value Theory and probabilistic worst-case execution time estimation, which are indispensable tools for the analysis of real-time systems. This short paper attempts to describe and bridge this gap.

The solvers of the simulators with underlying stiff equations usually come with a large number of parameters that need to be tuned, also known as configuration settings. The developers of such simulators provide, in most of the cases, a default set of values for such parameters which are adequate for large number of problems. However, there are cases where the convergence is not reached using the default parameters values. Often in these cases, the simulation time becomes very high and a manual stop is required. The complex correlation between parameters and their high impact on the algorithm convergence implies that manual fine tuning is very time consuming. The default parameters set is usually far from optimal with regard to simulation time. Therefore, for an application requiring a large number of simulations, it will be convenient to have a tool that minimizes the simulation time, while also maintaining the accuracy. This paper presents an analysis tool that can be used to speed-up the parameter tuning process by up to 10 times. This tool relies on Machine Learning algorithms and assures the convergence of the solver through active learning and adaptive verification.

International audience; The design of cyber-physical systems (CPSs) is facing the explosion of new functionalities requiring increased computation capacities and, thus, the introduction of multi-core processors. Moreover, some functionalities may impose precedence constraints between the programs implementing these new functionalities. While important effort has been dedicated to the scheduling of precedence constraints tasks on multi-core processors, existing work considers either partitioned scheduling for a single precedence graph defining precedence constraints between tasks, or global scheduling policies.In this paper, we consider partitioned scheduling for multiple precedence graphs defining precedence constraints between tasks. The variability of execution times and of communication times is described by probability distributions. We propose a new response time analysis over-performing existing ILP-based results. Thanks to its scalability, our solution is extendable to a probabilistic version and we validate it on a PX4 drone autopilot. Beside this autopilot for our experiments, we implemented a probabilistic extension of a multi-core processor simulator, SimSo. A priority assignment heuristic allowing parallel executions is also proposed. Thanks to its adaptation to partitioned scheduling, our heuristic has better performances than existing solutions and its performances are, also, compared against a genetic-based heuristic.

Work in progress ETFA20; International audience; Estimating bounds for the execution or response times of a task is a central concern for real-time designers. Several solutions exist, and probabilistic approaches estimate such bounds by building appropriate probability distributions. Those methods are safe, but they may be pessimistic and rely on strong hypothesis such as independence between tasks. The worst case execution times of tasks are hard to estimate because their measurements are usually disturbed by the system itelf. In general measures are done in isolation, however dependencies between tasks are rarely modeled in that case. By isolation, we mean that a task (or program) is executed without any type of interference coming from other executed tasks. In this paper we propose a statistical analysis of measured response times based on clustering analysis, i.e., building classes of response times that may identify executing modes for a given set of tasks. This work is a first step towards a multivariate analysis that may explicitly identify dependence structures.

High-performance automatic speech recognition (ASR) systems are regularly trained on tens of thousands of hours annotated speech (i.e. speech paired with correct transcripts). Collecting such amount of data is prohibitively costly if done manually (i.e. humans listening and transcribing audio clips). However, raw speech data (without transcripts) is widely available and easily collectable. This paper proposes an automatic method that uses approximate transcripts of raw speech and an already existing ASR system to generate annotations. The method is evaluated in terms of annotation efficiency (i.e. the percentage of the initial raw speech corpus for which it provides annotations) and in terms of data usefulness for further training ASR systems. We show that, although the method is able to produce less data than other methods, the ASR system retrained using the newly created dataset performs significantly better than the baseline. Furthermore, we report ASR results that are better by 17% to 25% than what was reported up to now on Romanian read and spontaneous speech.

This study compares the vocabulary of a single Economics text of almost 300,000 running words with the vocabulary of a corpus of similar length made up of a variety of academic texts. It was found that the general academic corpus used a very much larger vocabulary than the more focused Economics text. A small number of words that were closely related to the topic of the text occurred with very high frequency in the Economics text. The general academic corpus had a very large number of low frequency words. Beyond the words in West's General Service List and the University Word List, there was little overlap between the vocabulary of the two corpora. This indicates that as far as vocabulary is concerned, EAP courses that go beyond the high frequency academic vocabulary are of little value for learners with specific purposes. © 1994, Sage Publications. All rights reserved.

International audience; The design of embedded systems is facing the explosion of new functionalities requiring increased computation capacities and, thus, the introduction of multi-core processors. Moreover, some functionalities may impose precedence constraints between the programs implementing them. In this paper, we consider partitioned scheduling of tasks with precedence constraints defined by multiple Directed Acyclic Graphs (DAGs). The variability of execution and communication times is taken into account by describing them with probability distributions. Our probabilistic response time analysis is validated on random generated task sets and on a PX4 drone autopilot.

Designing and implementing artificial systems that can be interfaced with the human brain or that can provide computational ability akin to brain's processing information efficient style is crucial for understanding human brain fundamental operating principles and to unleashing the full potential of brain-inspired computing. As basic neural network components, responsible for information transfer between neurons, artificial synapses able to emulate analog biological synaptic behaviour are of particular interest. State of the art CMOS and memristor-based synapses suffer from scalability drawbacks (large energy consumption and area footprint), variability-induced instability, and are not bio-compatible. In this paper, we propose a generic Graphene Nanoribbon (GNR) based synapse structure and demonstrate that by changing GNR geometry and external bias voltages it can emulate different synaptic plasticity behaviours, i.e., Spike Timing Dependent Plasticity and LongTerm Depression and Potentiation, and that both excitatory and inhibitory synaptic behavior can be obtained with the same GNR geometry. To demonstrate biologically plausible operation, we make use of low voltage bias, i.e., 0.1V, 0.2 V, and consider inputs consistent with measured brain synapses data, i.e.,-50 mV to 50 mV pre-and post-synaptic spikes voltage range, and-60ms to 60 ms time range. The simulations indicate that by changing the GNR shape we can enrich the plasticity behaviour (potentially beyond the considered cases) and the plasticity change of 100% provided by natural synapses can be achieved. Our investigation clearly suggests that the proposed GNR synapse structure is a promising candidate for large-scale neuromorphic systems integration, which might potentially bring novel insight on brain neurophysiology, as it requires a small footprint, is energy effective, biocompatible, and versatile from the synaptic behaviour point of view.

International audience; The Worst-Case Execution Time (WCET) is a critical parameter describing the largest value for the execution time of programs. Even though such a parameter is very hard to attain, it is essential as part of guaranteeing a real-time system meets its timing requirements. The complexity of modern hardware has increased the challenges of statically analysing the WCET and reduced the reliability of purely measured the WCET. This has led to the emergence of probabilistic WCETs (pWCETs) analysis as a viable technique. The low probability of appearance of large execution times of a program has motivated the utilization of rare events theory like Extreme Value Theory (EVT). As pWCET estimation based on EVT has matured as a discipline, a number of open challenges have become apparent when applying the existing approaches. Our paper enumerates key challenges while establishing a state of the art of EVT-based pWCET estimation methods.

Designing and implementing artificial neuromorphic systems, which can provide biocompatible interfacing, or the human brain akin ability to efficiently process information, is paramount to the understanding of the human brain complex functionality. Energy-efficient, low-area, and biocompatible artificial neurons are key ubiquitous components of any large scale neural systems. Previous CMOS-based neurons implementations suffer from scalability drawbacks and cannot naturally mimic the analog behavior. Memristor and phase-changed neurons have variability-induced instability drawbacks, and usually rely on additional CMOS circuitry. However, graphene, despite its ballistic transport, inherently analog nature, and biocompatibility, which provide natural support for biologically plausible neuron implementations has only been considered for Boolean logic implementations. In this paper, we propose an ultra-compact, all graphene-based nonlinear Leaky Integrate-and-Fire spiking neuron. By means of SPICE simulations, we validate its basic functionality and investigate the output spikes response under stochastic noisy input spike trains with a variable firing rate, from 20 to 200 spikes per second. Simulation results indicate neuron robustness to noisy scenarios, and neuronal output firing regularity. The small area and the low energy consumption, due to 200 mV supply voltage operation, can benefit the implementation of large scale neural networks, and the biologically plausible operating conditions (e.g., 2 ms and 100 mV spike duration and amplitude), can promote the interfacebility of graphene-based artificial neurons with biological counterparts.

Nowadays, real-time embedded systems are facing concerns like power consumption and increased functionalities demand. Approaches based on Dynamic Voltage and Frequency Scaling (DVFS) reduce the energy consumed by processors while guaranteeing real-time constraints. In this paper, we present short-comings on existing models reducing energy consumption. Our experimental results clearly show that the execution time of tasks is not exclusively proportional to the processor speed. Thus, we believe that DVFS techniques could also be applied to other components like buses and memory. We discuss the applicability of a probabilistic Worst Case Execution Time (WCET) combined with DVFS techniques, and argue that by adopting a probabilistic frequency-aware model, we can (i) capture more detailed behaviors of tasks w.r.t. hardware frequencies and (ii) apply DVFS techniques to gain in energy consumption.

This paper presents some issues regarding vehicles energy efficiency which implies simultaneous approach on the dynamic and economic parameters. Nowadays, the interest in improving the dynamic performances of the vehicles together with reducing the fuel consumption has grown, issue that is directly related to energy efficiency. In the sense of those presented, the paper approaches the criteria of appreciation of the dynamics and economy used in the specialized literature. Also, there are defined and established some appreciation values criteria of energy efficiency of vehicles which addresses both aspects. The study is based on the experimental data obtained from the test of a gasoline powered passenger car, data acquired from the on-board computer by a specialized tester. The paper proposes a study method of dynamic performances of vehicles related to energy efficiency.

The active cycling (repetitive clamping) of power devices is a time consuming process. For this reason, a limited amount of reliability data is available and the manufacturers most often provide the lifetime parameters only for a few specific operating conditions. According to the well-known Coffin-Manson lifetime model, the key is the estimation of the maximum junction temperature swing. The paper proposes a methodology for estimation of the maximum junction temperature swings on different operating conditions, based on electro-thermal simulations. Furthermore, we propose an extension of the classical Coffin-Manson model so that it can be applied on different operating conditions and for a predefined failure criterion. The coefficients of the model are fitted based on experimental data. The leave-one-out and bootstrapping validation methods show a maximum relative error of 25%. The proposed application-aware lifetime model is robust and simple, having only a small number of coefficients to be fitted.

Kaldi NNET3 is at the moment the leading speech recognition toolkit on many well-known tasks such as LibriSpeech, TED-LIUM or TIMIT. Several versions of the time-delay neural network (TDNN) architecture were recently proposed, implemented and evaluated for acoustic modeling with Kaldi: plain TDNN, convolutional TDNN (CNN-TDNN), long short-term memory TDNN (TDNN-LSTM) and TDNN-LSTM with attention. To the best of our knowledge, this is the first paper to describe in-depth these various flavors of TDNN, providing details regarding the speech features used at input, constituent layers and their dimensionality, regularization techniques etc. The various acoustic models were evaluated in conjunction with n-gram and recurrent language models in an automatic speech recognition (ASR) experiment for the Romanian language. We report significantly better results over the previous ASR systems for the same Romanian ASR tasks.

Keyword spotting is the task of identifying words of interest in audio streams. The task has been approached in the past using HMM-based speech decoders and more recently with deep neural networks. This paper presents a two-fold approach in which the audio stream is first transcribed into words and then the words of interest are searched for in the transcription. We focus on the search part which is specifically tailored for the Romanian language. We propose two search methods for identifying words of interest in the transcriptions, namely, direct string search and dynamic time warping-based search. Both of them are applied on raw and, for better performance, on lemmatized transcriptions. The methods are evaluated in terms of F-measure. We report F-measures of up to 92% for the search algorithm solely and up to 89% for the full system: speech transcription followed by keyword spotting.

International audience; Technologies within embedded real-time systems are continuously evolving making such systems intelligent. This evolution has increased the interest in the data utilization while real-time constraints are considered. In this paper, we consider real-time constraints for programs communicating using a circular buffer communication paradigm. We propose a first result optimizing the buffer size. Our second contribution consists in providing an analytical characterization of the temporal validity and reachability properties of the data propagating along a functional chain. Last but not least, we propose a scheduling policy ensuring the consistency and the temporal validity of the used data.

International audience; The continuous integration of new functionality increases the complexity of embedded systems, while each functionality might impose precedence constraints between the programs fulfilling it. In addition, the prevalence of several processors may create the illusion of higher computation capacity easing the associated scheduling problem. However, this capacity is not exploitable in critical real time systems because of the increased variability of the execution times due to processor features designed to provide excellent average time behaviour and not necessarily ensuring small worst case bounds. This difficulty is added to the existence of scheduling anomalies when the systems are built on top of several processors. In this paper, we study the feasibility of independent tasks scheduled according to a given preemptive fixed-priority partitioned policy on identical processors. Each task is composed of several dependent subtasks related between them according to a directed acyclic graph (DAG). We provide a worst case response time analysis for DAG tasks when each sub-tasks have an individual priority level. This assumption allows to decrease the number of possible execution scenarios, making our analysis easier and less pessimistic.

This paper addresses the problem of building a speech recognition system attuned to the control of unmanned aerial vehicles (UAVs). Even though UAVs are becoming widespread, the task of creating voice interfaces for them is largely unaddressed. To this end, we introduce a multi-modal evaluation dataset for UAV control, consisting of spoken commands and associated images, which represent the visual context of what the UAV "sees" when the pilot utters the command. We provide baseline results and address two research directions: (i) how robust the language models are, given an incomplete list of commands at train time; (ii) how to incorporate visual information in the language model. We find that recurrent neural networks (RNNs) are a solution to both tasks: they can be successfully adapted using a small number of commands and they can be extended to use visual cues. Our results show that the image-based RNN outperforms its text-only counterpart even if the command-image training associations are automatically generated and inherently imperfect. The dataset and our code are available at http://kite.speed.pub.ro., 5 pages, accepted at Interspeech 2019

Deep learning techniques, requiring large amounts of training data, are currently state-of-the-art in automatic speech recognition (ASR). Corporate giants, such as Google or IBM, train English ASR systems on more than 100k hours of annotated speech, while research on under-resourced languages, such as Romanian, has to deal with as little as 300 hours. In this context, automatic annotation of speech corpora and unsupervised acoustic model training are promising directions to be explored to leverage the lack of data. This study describes the progress made by SpeeD laboratory in this research direction: using an already proven methodology, applying it on large scale (more than 700 hours of unlabeled speech) and analyzing in-depth the experimental results to identify potential future directions. Moreover, we present novel results on Romanian ASR: the methodology leads to a relative Word Error Rate (WER) improvement up to almost 10%.

Timely addressing baby cries is always a challenge for new parents. Our project aims to develop a baby cry recognition system, capable of distinguishing between different kinds of baby cries, in real-world conditions. This will inform parents of their specific baby need, while they learn to make the distinction for themselves. In this study, we describe a series of experiments designed to establish the accuracy of popular machine learning algorithms on the categorization of 7 types of baby cries. We tested the algorithms on our own baby cry database, SPLANN[1], containing over 13K baby cries, recorded in a neonatal hospital. We extract acoustic features, perform best feature selection and report increased classification accuracies, from a coin-toss rate of 14.2%.

Using the Activity Theory to Identify the Challenges of Designing eLearning Tools based on Machine Learning for Security Operations Centers Mihail CAZACU Economic Informatics Doctoral School, University of Economic Studies, Romania, Romana Square 6, Bucharest, Romania mihail.cazacu@gmail.com Maria-Iuliana DASC?LU Department of Engineering in Foreign Languages, Faculty of Engineering in Foreign Languages, University Politehnica of Bucharest, Splaiul Independentei, No 313, Bucharest, Romania maria.dascalu@upb.ro Constanta-Nicoleta BODEA Department of Economic Informatics and Cybernetics, Faculty of Economic Cybernetics, Statistics and Informatics, University of Economic Studies, Romana Square 6, Bucharest, Romania bodea@ase.ro Abstract There is a fast-growing requirement for setting up Security Operation Centers (SOCs), with qualified personnel, mainly due to the increase of demands to protect ITC systems from security breaches, data disruption or unauthorized usage. The 2018 Report of Privacy Rights Clearinghouse mentions that over 8,000 data breaches were reported since 2005, with more than 10 billion records affected. And according to the 2017 study of IBM Security and Ponemon Institute, the average cost of a data breach exceed 3.6 million US dollars. SOCs have the mission to run in this "arms race" against cyber attackers (criminals, spies, terrorists, activists) and to be economically viable, as a profit or a cost center. Development of e-learning tools for continuous enhancing of the professional competences of the SOC's personnel is critical for the successful operation of SOCs. Recent studies have applied the framework of the Activity Theory in order to identify the conflicting priorities which need to be handled by different members of SOCs and have suggested ways to mitigate the risks. While automating mundane tasks is one solution, the issue of automating the automation process itself through Machine Learning, especially in the e-learning activities performed inside SOCs was not often addressed. The paper aims to present the challenges of applying the framework of the Activity Theory in designing e-learning tools based on machine learning methods for SOCs. Some well-established Open Source security tools and machine learning packages will be evaluated for their suitability for developing e-learning tools.

The large majority of the current cybersecurity technologies are identifying cyberattacks based on predefined signatures, rules, filters or scenarios such as virus or malware definitions, firewall rules, intrusion prevention filters or SIEM playbooks. In recent years, these technologies have achieved only a small reach towards addressing the new challenges and the pace of their development is bowing in front of the much faster evolution of the threats. This exponential increase in the number and complexity of cyberattacks opened the door to the utilization of artificial intelligence algorithms within the protection technologies primarily for a couple of reasons: to identify deviations from trained regular behavior and to analyze large volumes of data for patterns. Following this cycle, in turn, cyberattacks started themselves to employ artificial intelligence to automate labor-intensive tasks like social media and other public information analysis to prioritize targets, to evade current techniques, to generate domains or to perform human-like operations. These new trends started to challenge the current static-rule-based approach largely used today in the cybersecurity ecosystem and at the same time to provide the opportunity to adapt the current technologies and sharing platforms to respond to these dynamic threats. The learning processes needs modifications as well by employing mechanisms that will allow to deliver knowledge faster and more specialized. These developments in the cybersecurity realm led not only to the adaptation of the existing technologies but also to the creation of new trends, new technologies and new products that could offer alternative protection against this fast changing threat landscape.