Dr. Korhan Cengiz

Trakya University, Edirne, Turkey

Novel WSN Protocols for Health Care and Critical Applications

The reduction of energy consumption has become a key research area for the information and communication technology (ICT) industry, due to economic, environmental, and marketing reasons. While the environmental direction aims at minimization of greenhouse gas emissions by enforcing the usage of renewable energy in the ICT industry, economical and marketing directions lead researchers to design low-power components or develop and enhance energy-saving protocols without an impact on the level of the performance. With the steady increase in the cost of energy, the expanding number of energy-hungry components and widespread usage of ICT industry, most of the protocols that have become an integral part of our lives but are yet developed without any energy constraints in mind in the past will need to be restructured or developed again. For this reason, researchers are studying on all layers of the Internet protocol stack to develop energy-efficient protocols and algorithms. This keynote lecture reviews recent approaches for energy efficiency studies for each layer in the Internet protocol stack from the physical layer to the application layer and also especially for WSNs. It is expected that with the deployment of current research output, the studies performed at each layer will result in significant energy savings for the ICT industry which in turn will have a positive impact on our lives for their economic and environmental results.

Prof. Timo Jämsä

University of Oulu, Oulu, Finland

Analyzing physical activity data collected with accelerometers

Wearable accelerometer-based devices have become very popular in consumer applications, and they are also used widely in research. There is a wide range of algorithms, activity parameters and applications for monitoring daily physical activity (PA). However, interpretation of accelerometric data is strongly dependent on the purpose and health outcome in question, and there is not a single approach. Traditional solutions include counting the number of daily steps or estimating the intensity of PA as metabolic equivalents (MET). More recently, machine learning methodologies have also been applied for assessing and classifying PA. This presentation overviews some approaches and algorithms for assessing physical activities and sedentary behavior, applying data collected with an accelerometer in daily life.

Prof. Martin Lukac

Nazarbayev University, Kazakhstan

Optimization of Convolutional Neural Networks

Neural networks take a lot of resources to train and do inference. The general trend is to train multi-class and even multi-task end-to-end training networks in order to minimize the human effort and maximize the usability of the resulting system. However and in particular Convolutional Neural Networks are very resource consuming. In this work we look at two aspects of reducing the computational requirements of CNNs. First, we analyze some of the off-the-shelf available CNNs for object detection and determine the amount of computational overhead present by an adaptive pruning method. The proposed method is analogous to some previously introduced methods to study CNNs. It starts from a larger pretrained network and prune it until it behaves as a binary classifier. The results show that a) it is possible to prune up to 50% of convolutional filters without the loss of classification accuracy for one class of objects and b) a well controlled process of pruning does not affect the accuracy of the target class to be classified. In addition, we study the capacity of binary networks. Binary models were shown to significantly reduce the memory and computational resources, at the cost of a lower accuracy. The aim of this paper is to provide the further exploration of the binarization effect on the model capacity. To achieve even more control over the binarization process, we binarize the input images, which was not performed previously, and combine features of several binary networks to perform a classification task. Each image is fitted to the separate binary CNN model. Testing is performed by assembling separate models to predict the class label. The results show that while for MNIST the accuracy is very close to full precision counterpart, for the more complex dataset, CIFAR-10, the binarization and the representational power of CNNs are strongly affected.

Prof. Dmitry A. Zaitsev

Odessa State Environmental University, Ukraine
Supercomputación Castilla y León, Spain

Infinite Petri Nets for Cybersecurity of Intelligent Networks, Grid, and Clouds

Correctness of networking protocols represents the principal requirement of cybersecurity. A new class of infinite Petri nets has been introduced and studied for modeling modern networks, clusters, computing grids, and clouds. Parametric description of infinite Petri nets, parametric representation of infinite linear systems for the calculation of place/transition invariants, and solving them in parametric form allowed the invariance proof for infinite Petri net models. Some additional analysis techniques based on graphs of transmissions and blockings are presented. Complex deadlocks has been revealed and classifies as: a loop of blockings; a chain of blockings ended on an already blocked vertex; because of isolation by neighboring blocked devices. Results are generalized on multidimensional structures such as hypercube and hypertorus which play a key role in communication subsystems of supercomputers, clusters, and networks on chip. Generators of Petri net models has been developed and put on GitHub for public use. As a result of complex deadlocks disclosure, a possibility of network blocking via ill-intentioned traffic has been revealed.