Marija Punt

Primene veštačke inteligencije u obrazovanju danas su brojne, među kojima se kao značajna izdvaja predikcija akademskog uspeha studenata metodama mašinskog učenja. Cilj ovog rada je identifikacija i analiza trendova u primeni metoda mašinskog učenja u predikciji akademskog uspeha studenata u visokom obrazovanju u poslednjih pet godina. Sistematski pregled literature u ovoj oblasti sproveden je na osnovu unapred definisanih kriterijuma, a za obezbeđivanje transparentnosti i nepristrasnosti korišćene su smernice ukazane PRISMA protokolom za sistematske preglede literature. U analizu je uključeno 46 studija. Rezultati su pokazali da se najčešće u predikciji akademskog uspeha implementiraju ansambl i klasični modeli, a da je izuzetno učestalo eksperimentisanje sa hibridnim modelima. Performanse modela su evaluirane korišćenjem metrika od kojih su najčešće tačnost i površina ispod ROC krive. Podaci koje ove studije koriste su prevashodno akademski i demografski, uz porast upotrebe podataka sa LMS platformi.

This paper proposes a novel approach to speech-like signal design for reliable real-time transmission of encrypted data over voice channels. The structure of the digitally encrypted information-carrying segment relies on the known concept of the OFDM/QPSK modulation. The segment is extended with time-synchronization sequences, amplitude variations and silence intervals, all necessary for successful passing through realistic channels. Phase distortion has been identified as a significant source of error and inspired the proposal of a phase-shift compensation and a time-and-phase fine-tuning algorithms. Finally, this paper outlines a complete secure real-time communication system. Transmission quality of the resulting speech-like scheme is first evaluated on the AMR codec at various compression rates, followed by the analysis over realistic voice channels. The proposed secure DoV system achieves a mean BER of $2.45\cdot 10^{-2}$, $1.11\cdot 10^{-2}$ and $4.83\cdot 10^{-3}$ over Signal, Telegram and 3G cellular network voice channels, respectively.

The Internet of Things (IoT) connects a wide range of sensors and devices in environments that are often dynamic and resource-constrained, where efficient distributed solutions are essential for ensuring robust and scalable operation. This article presents a novel adaptive consensus algorithm, tailored for distributed signal processing in heterogeneous sensor networks, with a focus on distributed estimation and target tracking. The algorithm addresses the challenge posed by networks where intelligent sensors have limited sensing, computation, and communication capabilities, resulting in diverse quality of locally available information, and leading to neighbor-based information exchanges. It employs asynchronous gossip protocols to randomly exchange information between nodes, ensuring robustness to time synchronization and network topology uncertainties, while limiting computational and communication costs. The adaptation mechanism operates in two complementary ways. First, we account for variations in the quality of local processing results, ensuring that asymptotic behavior of the consensus scheme accurately reflects this diversity. Second, we introduce a novel adaptation of the rates at which nodes initiate communication, using the available local information. This enables fast information dissemination and provides a solution that is both effective and efficient. We show that, under appropriate network connectivity assumptions, the results obtained by the algorithm converge to the desired asymptotic values in the mean square sense. Numerical simulations demonstrate the algorithm’s properties and effectiveness, particularly in modeling visual surveillance networks where fixed cameras are augmented by moving drones to extend the coverage area.

The core objective is to securely transmit digital data over widely accessible voice channels like VoIP or GSM networks. Encryption at the highest security level is applied digitally, but a challenge lies in selecting a modulation technique that effectively converts encrypted digital data into analogue signals capable of navigating the obstacles inherent in voice channels. These channels are lossy and optimized for speech transmission, employing sophisticated signal processing techniques such as Voice Activity Detection (VAD) to maintain speech quality and suppress noise. This paper introduces a novel system design and approach for transmitting Data over Voice (DoV) channels using Offset Quadrature Amplitude Modulation (OQAM) combined with Filter Bank Multicarrier (FBMC) modulation. FBMC/OQAM generates speech-like waveforms that can evade detection by non-speech signal detection mechanisms inherent in voice channels, marking its debut in this domain. In order to demonstrate the feasibility of the proposed system, we have simulated the distortion of FBMC/OQAM signals when processed by an AMR-NB codec using MATLAB®. Subsequently, the performance of the proposed modulation technique is evaluated across various real voice channels, considering actual time and phase deformations, real-time voice transmission bit-rate constraints and possible bit errors, which all demand specific waveform structures for synchronization and algorithms for correct demodulation and decoding at the receiver. Evaluation is done for waveforms that were sent through GSM and VoIP channels. The obtained results demonstrate a favorable bit error rate (BER) on the 3G channel compared to existing methods, such as conventional Orthogonal Frequency Division Modulation (OFDM), thereby proving that FBMC/OQAM is suitable for secure DoV transmission. Finally, the decoded speech from all channels is assessed using Perceptual Evaluation of Speech Quality (PESQ) metric.

The goal of this paper is to inform the machine learning community of our results obtained during the development of a system for the assessment of the postoperative lung function of patients suffering from lung cancer. The system is based on a new multilayer regression meta-model, which predicts individual postoperative forced expiratory volume in 1 second (poFEV1) for each patient based on preoperative measurements. The proposed regression models are especially trained to predict this key indicator for the 1st, 4th, and 7th day after surgery. Based on our knowledge, this is the first attempt to obtain poFEV1 in the most critical postoperative period of the first seven days. The high accuracy of the proposed predictive meta-model allows surgeons a number of key insights, starting with whether the patient is suitable for surgical intervention, and ending with the preparation of individualized postoperative treatment. It should be noted that the existing, widely used predictive models, based on functional segments (FC), Juhl-Forst, and Nakahara formulas, give two to three times worse results compared to the proposed new regression meta-model. Based on the SHapley Additive explanations (SHAP) value of the trained meta-model, it is possible to obtain a complete picture of the partial effects of each prognostic factor for each patient preoperatively on the outcome of the surgical intervention. In addition, the global model interpretation by SHAP values reveals some new interdependencies that were not known in medical circles until now. For instance, the influence of age and biomass index on the condition of the patient on the first day after surgery, or the constant significant influence of muscular support for inhalation in the entire seven-day follow-up period.

Ensemble clustering, a paradigm that deals with combining the results of multiple clusterings into a single solution, has been widely studied in recent years. The goal of this study is to propose a novel distributed ensemble clustering method that is applicable for use in networked multi-agent systems. The adopted setting supports both object-distributed and feature-distributed clusterings. It is not limited to specific types of algorithms used for obtaining local data labels. The method assumes local processing of local data by the individual agents and neighbor-wise communication of the processed information between the neighboring agents in the network. Using the proposed communication scheme, all agents are able to achieve reliable global results in a fully decentralized way. The network communication design is based on the multi-agent consensus averaging algorithm applied to clustering similarity matrices. It provably results in the fastest convergence to the desired asymptotic values. Several simulation examples illustrate the performance of the proposed distributed solution in different scenarios, including diverse datasets, networks, and applications within the multimedia domain. They show that the obtained performance is very close to that of the corresponding centralized solution.

A growing number of research papers shed light on automated machine learning (AutoML) frameworks, which are becoming a promising solution for building complex machine learning models without human expertise and assistance. The key challenge in enabling AutoML frameworks to build an efficient model for anomaly detection tasks is to determine the best underlying model for a given task and optimization metric. The meta-learning approaches based on a set of meta-features that describes data properties can enable efficient model selection in AutoML frameworks. The existing meta-learning approaches based on statistical and information-theoretic meta-features require large amounts of data and computational resources to extract data properties. This paper proposes a novel set of meta-features for model selection in anomaly detection tasks based on domain-specific properties of data which overcomes the shortcomings of existing meta-features by introducing simple but effective meta-features that can be efficiently extracted or estimated by using a low amount of data. Experiments with 63 datasets from different repositories with varying schemas show that the proposed set of meta-features achieves an accuracy of 87% for model selection, while the achieved accuracy for simple meta-features is 74%, for statistical meta-features 68%, for information theory meta-feature 70%, and for a comprehensive set of meta-features by pyMFE 73%. This demonstrates that the proposed set can be adopted by AutoML frameworks across a diverse range of domains.

In this paper an implementation of an application for detecting and tracking faces in real-time using the Viola-Jones algorithm and OpenCV library has been presented. Also, using the developed application it is possible to recover from failure during face tracking. Video content is obtained from a file or webcam.

In this paper a cloud-based TV-centric gaming architecture is proposed, enabling the creation of games played amongst friends seated in geographically separated living rooms. In each living room a television is used as a public game display and mobile devices are used as game controllers. An example of a multi-player board game using the proposed architecture was developed and validated using tests measuring the communication performance between server and client.