Young Researchers Day
Agenda

Tensor Decompositions for Big Data Analytics: Blessing of Dimensionality

Prof. Dr Danilo P. Mandić

Abstract. Multidimensional arrays, also known as tensors, are data structures which are routinely recorded from many modern sensors such as RGB images, multichannel Electroencephalogram, or even communication network traffic data. Yet, standard algorithms for the analysis of such data are typicaly based on matrices, and assume a "flat-view" stracture of the multidimenional world. We show that by tensor representations and the associated tensor decompositions (TD) are an excellent match for the high dimensionality of modern data, whereby the associated super-compression ability of tensor decompositions not only makes the problem computationally tractable, but also preserves physical meaning and interpretability of the so derived latent components in data. A link between tensors and deep learning is also highlighted, with a special emphasis on the link with tensor networks, an arrangement of very high order tensors into lower order cores which are easier to compute on and can be stored in a distributed way or even on different computers. Applications in computer vision, biosignal processing and multimodal learning support the analysis.Multidimensional arrays, also known as tensors, are data structures which are routinely recorded from many modern sensors such as RGB images, multichannel Electroencephalogram, or even communication network traffic data. Yet, standard algorithms for the analysis of such data are typicaly based on matrices, and assume a "flat-view" stracture of the multidimenional world. We show that by tensor representations and the associated tensor decompositions (TD) are an excellent match for the high dimensionality of modern data, whereby the associated super-compression ability of tensor decompositions not only makes the problem computationally tractable, but also preserves physical meaning and interpretability of the so derived latent components in data. A link between tensors and deep learning is also highlighted, with a special emphasis on the link with tensor networks, an arrangement of very high order tensors into lower order cores which are easier to compute on and can be stored in a distributed way or even on different computers. Applications in computer vision, biosignal processing and multimodal learning support the analysis.

5G: Key Enabler for Internet of Things, Smart Cities and Industry 4.0

Aleksandar Mastilović, MSc

Abstract. Modern challenges in telecommunications are defined with new applications as Internet of Things set of services, Smart Cities, Smart Grid and many other smart solutions but also there is a need for additional automatization of the industry processes, what is defined as Industry 4.0. All these appolications and many others are based on the system of massive number of sensors. Sensors collect the data and send them to nodes for storage and analytics. Current achitecture of telecommunications systems including modern mobile/cellular system 4G (LTE) are not ready to serve for massive and critical Machine-to-Machine communications. 5G is a key enabler for new generation of services and 5G has strictly defined technical parameters and different architecture to be able to reach predefined performances and realibility. New algorithms and new approach for mobile systems will be presented.

Semantics for Interoperable IoT Applications

Dr Darko Aničić

Abstract. The Internet of Things (IoT) is fragmented today as it lacks the interoperability across diverse platforms, ecosystems, and domains. As a result, IoT devices and applications are not interoperable. They operate with data silos, and the development of new business applications are faced with high costs and limited market potential. The World Wide Web Consortium (W3C) has recently started a standardization activity called Web of Things (WoT) to encounter this problem. The upcoming standard introduces a Web layer to IoT with the goal of reducing IoT fragmentation. WoT provides rich metadata that describes capabilities of Things, data they exchange, communication and security related information, as well as an API needed to interact with Things. In order to enable applications and platforms to share the same meaning when they exchange data, a semantic interoperability layer at the application level is needed. Applications that are tightly coupled to IoT devices will make up a smaller fraction of delivered value of connected Things over time. There will be more value in interoperable applications that can orchestrate behavior across diverse devices and device ecosystems. This is the reason why iot.schema.org has recently started to work on semantic models that describe common affordances of connected Things (e.g., What does a Thing do? What can I control with it?). Further on, iot.schema.org provides semantics needed for interactions with connected Things from different device ecosystems, i.e., Things that use similar protocols but diverse data models. In this talk, an overview of both activities, W3C WoT and iot.schema.org, will be provided.

Semantic Segmentation of Ranging Sensor Data

Vladimir Lekić, MSc

Abstract. Ranging sensors, such as radar and lidar, on-board the vehicle are considered to be very robust under changing environment conditions. Largely owing to this reputation, they have found a broad applicability in driver assistance, and consequently in autonomous driving systems. On the other hand, they lack precision. This makes classification tasks of the measurement data rather difficult. A method for semantic segmentation of the ranging sensors data using the generative adversarial networks will be presented. Utilizing the fully unsupervised learning algorithm sensor data is converted into artificial, camera-like, environmental images that are further used as input for semantic image segmentation algorithms.

48V System Made by Mercedes

Kemal Hajvazović

Abstract. In modern vehicles the number of the electrical consumers and the energy demand is constantly increasing. Equipment such as fast seat heaters, armrest heaters, massage function or Thermo-Cupholder improve comfort, but at the same time put additional performance requirements on the electrical system. The base of the energy supply for today’s passenger cars is the 12V electrical system. This system is established and did not change significantly for decades. The operating voltage range is from 11 to 15 volts. This limited voltage range leads, with increasing power requirement, to a steady increase of the occurring currents. Higher currents lead to bigger lead acid batteries and heavier vehicle harness, which directly influences the fuel economy. Therefore present 12V system does not offer enough reserves for future applications. In order to tackle this problem Mercedes-Benz uses a combination of 48 V and 12 V electrical systems in the new S-Class. The 48 V electrical system solves the above stated conflict without the disadvantages of a costly high-voltage electrical system. Additionally it brings many of the well-known hybrid functions and opens potentials for the reduction of the fuel consumption. This presentation will show you current state of the implementation of the 48V system in the new S-Class and highlight future development directions.

HYLO – HYbrid LOgarithmic Multiplier

Ratko Pilipović, ma

Abstract. Much attention is devoted to the design of multipliers as one of the most important arithmetic circuit. By allowing small but controllable computational error a efficient design of multiplier can be achieved. Although much work is done, leading inexact multiplier design strategies, truncated and logarithm multipliers, fail to meet requirements for chip area efficiency and computational error. Presentation introduces a novel design paradigm for inexact multiplier which combines the advantages of leading design strategies. On the top level, proposed strategy employs decomposed multiplier, where the least significant partial product is obtained by utilizing logarithm multiplier. This allows design of power efficient multiplier which characterize with small error performance. Error performance of proposed method is evaluated by conducting error analysis and comparing it with state-of-the-art multipliers. The results show potential for further development and application of a multiplier in neuro-computing hardware.

Design and Characterization of an Inductor and a Low-Noise Amplifier in Monolithic Integrated Circuit Technology for Wideband Operation

Dr Aleksandar Pajkanović

Abstract. A passive inductive component and a low-noise amplifier are designed, fabricated in standard monolithic CMOS technology and characterized, both intended for wideband operation. For the design of the inductor, meander topology is chosen. Along with the integrated circuit design tools, electromagnetic field simulator is used. Besides the standard parameter characterization, special attention is dedicated to the analysis of process and temperature variations. Furthermore, mechanical characterization of the material that comprises the protection layer has been undertaken. Low-noise amplifier is designed as the first stage of an ultra wideband receiver and the results show that the circuit is successfully designed.

Expressive Speech Synthesis Based on Deep Neural Networks

Tijana Delić, ma

Abstract. Studies have shown that people already perceive the interaction with computers, robots and media in the same way as they perceive social communication with other people. For that reason it is critical for a high-quality text-to-speech system (TTS) to sound as human-like as possible. Even though synthetic voices are now intelligible and natural enough, they lack expression. Recent development of software and hardware increased the popularity of using artificial neural networks in many areas, including speech processing. One of the advantages of such methods in speech synthesis is the possibility to easily change speaker characteristics and speaking styles. The main approaches in the area will be presented, putting the accent on having a small amount of data which is typical for scarce languages.

ASR Recognition of Whispered Speech Based on Hidden Markov Models

Jovan Galić, ma

Abstract. Due to the lack of sufficient amount of whisper data for training, whispered speech recognition is a serious challenge for state-of-the-art Automatic Speech Recognition (ASR) systems. Because of great acoustic mismatch between neutral and whispered speech, ASR systems are faced with significant drop of performance when applied to whisper. In this presentation, an analysis of neutral and whispered speech recognition based on Hidden Markov Models (HMM) framework, in a Speaker Dependent (SD) and Speaker Independent (SI) cases will be given. Special attention is paid to the neutral-trained recognition of whispered speech (N/W scenario). As well, advantages and disadvantages related to recognition based on SVM (Support Vector Machines), ANN (Artificial Neural Networks) and DTW (Dynamic Time Warping) techniques will be discussed.

Algorithmic Optimization of Transmission Parameters in IEEE 802.11 WLAN

Boris Malčić, ma

Abstract. The main purpose of this research is to achieve a transmission optimization in real conditions of a dynamic environment. The optimization is based on an idea of mathematical modeling of data flow changes and spatial scheduling of clients in IEEE 802.11 WLAN. In order of existing state optimization in terms of reducing channel interference with energy-efficient work of WAPs in WLAN, new algorithms are proposed. Both algorithms (DCA and TCRFFA) make a unique proposal of the algorithm for the optimization. Also, an algorithm based solution for smart WLAN antenna is modelled in Matlab with a simple option of radio-localization in WLAN. Performed simulations have shown, that proposed solutions could lead to a decrease in EVM, improvements in QoS and transmissions at higher bit rates.

Data Aggregation for Advanced Communications in IEEE802.15.4e TSCH Wireless Sensor Networks

Dragan Vasiljević, MSc

Abstract. Wireless Sensor Networks – WSNs have a wide range of applications. One of them are industrial applications. This field is of particular interest because it requires WSNs to have a high performance in terms of reliability, throughput and data transmission delay. The usage of frequency hopping technique, which increases resilience to interference and multi path fading, enables IEEE802.15.4e TSCH protocol to have potential to satisfy the criteria of industrial applications. OpenWSN is an open software platform which implements IEEE802.15.4e TSCH protocol together with other standardized protocols for WSN thus making entire protocol stack for WSN. Previous researches using OpenWSN platform showed that the key factor for performance of IEEE802.15.4e TSCH WSNs is timeslot scheduling at MAC (Medium Access Control) layer. For a certain scheduling, additional performance increase can be achieved using the technique of data aggregation at WSN nodes. This technique is applicable in conditions where an average packet size in WSN is significantly smaller than maximum possible packet size. Data aggregation technique provides higher throughput, congestion avoidance and decreases packet losses.