UNIVERSITY OF CRETE Computer Science Department The Discreet Charm of Networking: From the Internet to the Brain Professor Maria Papadopouli Department of Computer Science - University of Crete Institute of Computer Science - Foundation for Research & Technology – Hellas (FORTH) http://www.ics.forth.gr/tnl http://www.ics.forth.gr/mobile http://www.ics.forth.gr/neuronxnet
Telecommunications & Networks Laboratory ry • Measuring & Mitigating Internet Routing Attacks 3 ERC grants (1 StG & 2 PoC) • Quality of Experience in the Internet • Internet of Things • Communication optimization • Secure over-the-air programming as a service • Remote compilation & monitoring • Healthcare & environmental applications • Supply chain traceability • Smart City Backbone development in real environments: Heraklion & Larnaca • Brain Network Analysis
Some Recent Research Activities • Internet of Things, Sensor Networks • Networks & Mobile Computing in Health-Care • Network Paradigms for better Quality of Experience in Video Streaming & VR/AR • Brain networks Analysis & Modeling • Complex real-world networks • Interdisciplinarity • Use of sophisticated machine-learning & statistical analysis algorithms (e.g., RQA, GANs)
1) New Network Paradigms for Improving QoE for VR/AR 2) GestureKeeper Gesture Recognition for Controlling Devices in IoT Environments Vasileios Sideridis , Andrew Zacharakis, George Tzagkarakis, Maria Papadopouli. 27th European Signal Processing Conference EUSIPCO 2019 3) DysLexML Robust & accurate screening tool for dyslexia with data augmentation using GANs Thomais Asvestopoulou , Victoria Manousaki, Antonis Psistakis, Erjona Nikolli, Vassilios Andreadakis, Ioannis M. Aslanides, Yannis Pantazis, Ioannis Smyrnakis, Maria Papadopouli. 19th IEEE International Conference on BioInformatics & BioEngineering BIBE 2019 4) neuronXnet Brain networks analysis & Modeling Andreas Zacharakis, Manthos Kampourakis, Orestis Mousouros, Ganna Palagina, Jochen Meyer, Stelios Manolis Smirnakis, Ioannis Smirnakis, Maria Papadopouli. Functional Network Connectivity Analysis in Absence Epilepsy Using Stargazer Mice. IEEE 19th International Conference on BioInformatics & BioEngineering, BIBE 2019
Network Challenges in AR/VR: Severe Latency Constraints The 1ms round-trip delay is still challenging due to the backhaul bottleneck that cloud servers face in the Internet Edge Computing • A three-tier architecture of cloud servers, edge nodes & end user devices or IoT nodes • Edge: small cells with ample resources of storage, computation and communication • Prediction of full-body movement • Rendering • Prefetching of relevant content • Takes advantage of the spatial locality of information to serve requests locally, avoiding directing them to the cloud → Significant improvement of the user experience and cost effectiveness
Approaches for Gesture Recognition f for Supporting an Io IoT Envir ironment 1) Camera-based • High recognition accuracy • High computational cost & sensitivity in environmental conditions 2) Sensor-based • Practical considerations (e.g., smaller sensor size, efficient, low cost) • Energy constrains 3) Wireless Access-Point based • Practical considerations. No intervention • Sensitivity in dynamic environmental conditions 8
GestureKeeper Long-term objective: Allow user to control devices through hand gestures Hand-gesture identification & recognition based on wearable inertial measurements unit (IMU ) Identify the start of gesture by exploiting underlying dynamics of collected time-series • First automatic hand-gesture identification system based only on accelerometers Recognize accurately a dictionary of 12 hand-gestures • Wearable sensor sends periodically collected measurements to server • Server performs gesture identification & recognition 9
GestureKeeper: Aim ims to recogniz ize each perf rformed gesture Shimmer 3 • Includes 3-axis accelerometer, gyroscope, magnetometer • Small detection range for the accelerometer • Sampling frequency: 50 Hz • Data streaming & logging 10
Id Identification: Aim ims to o id identify fy tim time win indows con ontain inin ing gestures & & act activit ities of of dail aily livi living (A (ADLs) • Feature vectors produced from transformations on acceleration data • Classifies the data in two classes: gestures & ADL 11
Recurrence Quantifi fication Analysis (R (RQA) • Powerful tool that uses theory of non-linear dynamics based on the topological analysis of the phase space of the underlying dynamics • Enables the understanding of the behavior of a complex dynamic system • Does not make any assumption about the model that governs the system or the data • Can handle short time-series, non-stationary data • Is robust to outliers 12
Towards a robust and accurate screening tool for dyslexia with data augmentation using GANs Thomais Asvestopoulou 1,2 , Victoria Manousaki 1,2 , Antonis Psistakis 1,2 , Erjona Nikolli 1,2 , Vassilios Andreadakis 3 , Ioannis M. Aslanides 4 , Yannis Panatazis 5 , Ioannis Smyrnakis 2,3 and Maria Papadopouli 1,2 1 Department of Computer Science, University of Crete, Heraklion, Greece 2 Institute of Computer Science, Foundation for Research and Technology-Hellas, Heraklion, Greece 3 Optotect Ltd., Heraklion, Greece 4 Emmetropia Eye Institute, Heraklion, Greece, 5 Institute of Applied and Computational Mathematics, Heraklion, Greece 13
Dyslexia is a reading disorder characterized by difficulties with accurate and/or fluent word recognition and by poor spelling and decoding abilities. G. R. Lyon, S. E. Shaywitz, B. A. Shaywitz, “A definition of dyslexia”, Annals of Dyslexia , vol. 53, pp 1-14, 2003 Early intervention can be effective in alleviating the symptoms of the disability. However, screening large populations of children is rather time consuming and expensive . Is it possible to build a screening tool that can reliably identify dyslexic readers by analyzing oculomotor patterns collected during reading and is robust under noise ? Could we address the lack of large-size datasets? 14
Typical Reader Dyslexic Reader Dyslexic readers manifest: • longer and more frequent fixations • shorter saccade lengths • more backward refixations 16
DysLexML Cross- Feature fixation validation extraction points training set test set LASSO regression Feature selection Dominant Extracting features dominant selection features Build test classification classifier classifier accuracy 17
Classification Algorithms 1. Support Vector Machines (linear, Gaussian, polynomial) 2. Na ϊ ve Bayes (continuous values, Gaussian distribution) 3. K-means (variation to fit classification, aiming to catch possible sub-classes) Leave One Out Cross Validation (LOOCV) for calculating the mean accuracy 18
Dominant Features • Number of fixations • Number of short forward movements • Fixation median duration • Median length of medium forward movements • Number of multiply fixated words • Age 19
1 st field study 69 participants custom made age span 8.5- use of chin rest (39 typical, 32 dyslexic) eye-tracker 12.5 years old Recording images up to 60Hz, resolution of 1600 x 1200 pixels 2nd field study 152 participants Tobii 4C Non-invasive age span 7-17 (80 typical, 72 dyslexic) eye-tracker procedure years old Recording images up to 90Hz (no chin rest was used) 20 Easy text: 143 words, mostly of 1 or 2 syllables All participants were native Greek speakers Difficult text: 181 words, many multi-syllable
DysLexML: Findings • Dominant features are interpretable and accepted by the community • For noise levels smaller than 30 pixels, system performs robustly Synthetic data generation for feeding more data-greedy & sophisticated classification methods to increase the accuracy This work sets the basis for developing screening systems that can reach larger more diverse populations, in less controlled environments , for early intervention and potentially larger social impact 21
Brain Networks Analysis & Modeling • Network analysis of visual cortex during learning • Encoding and adaptation in primary visual cortex • Neural networks in neurological diseases • Deciphering the role of neuronal activity patterns in epilepsy • Developing biomarkers for the progression of Alzheimer’s Disease Collaborator /co-Investigator Prof Stelios Smirnakis Department of Neurology Harvard Medical School 22
https://www.ics.forth.gr/tnl/
Recommend
More recommend