Keynote Speakers
Neuroinformatics - 2016
Monday, April 25 11.20 – 13.00
Lecture-hall Алексеевский зал
Chair: Prof. DUNIN-BARKOWSKI WITALI
ChemRar High-Tech Center, Khimki, Moscow region,
Roadmap NeuroNet National technology initiative
2. SHUMSKY S.A.
The Moscow Institute of Physics and Technology (State University)
Deep learning revolution: state-of-the-art and prospects of machine intelligence
Survey of the latest progress of the current "deep learning revolution", its roots, driving forces, achievements and challenges of creating an artificial intelligence.
Monday, April 25 14.00 – 16.30
Lecture-hall Алексеевский зал
Chair: Prof. SHUMSKIY SERGEY
1University of Leicester, Great Britain
2Toyota Research Institute of North America, USA
Paradoxes of randomized choice of basis
In this work we discuss the problem of selecting suitable approximators from families of parameterized elementary functions that are known to be dense in a Hilbert space of functions. We show that if parameters of such elementary functions are chosen at random then one may observe exponential growth of the number of terms needed to approximate the function and/or extreme sensitivity of the outcome of the approximation to parameters. Implications of the analysis are illustrated with examples.
4. GOLOVKO V.A.
Brest State Technical University, Belarus
Deep learning: theory and applications
This talk provides an overview and new trends in the evolution of deep neural networks. The basic techniques of deep learning based on restricted Boltzmann machine (RBM), auto-associative approach and stochastic gradient descent (SGD) with rectified linear unit (ReLU) activation function are discussed. In this work a new technique called “REBA” for training of deep neural network, based on restricted Boltzmann machine is proposed. We have shown that classical equations for RBM training are special case of proposed technique. It is proved, that maximization of the log-likelihood input data distribution P(x) of restricted Boltzmann machine is equivalent to minimizing the reconstruction mean squared error in RBM. The applications of deep neural networks for data compression, visualization and recognition are discussed.
5. STANKEVICH L.A., SONKIN K.M., SHEMYAKINA N.V., NAGORNOVA ZH.V., KHOMENKO YU.V.
1Peter the Great St. Petersburg Polytechnic University
2Sechenov Institute of Evolutionary Physiology and Biochemistry of the Russian Academy of Sciences, St.Petersburg
3International Scientific Center "Arktika", FEB RAS, Magadan
4Bekhtereva Human Brain Institute RAS, St.Petersburg
Non-invasive brain-computer interfaces. Classification of EEG patterns of imaginary movements
In this study directions of development of brain-computer interfaces (BCI) are discussed. Special attention is paid to noninvasive BCI created on the basis of electroencephalographic (EEG) signals analysis and recognition. One of the development directions is connected with the elaboration of suitable for BCI means of EEG-pattern classification of imaginary motor commands related to different parts movements. It is shown that the development of such means is an actual task because it can significantly improve the efficiency of modern noninvasive BCIs. Problems of working out of algorithms and programs of EEG signal preprocessing and generated EEG-pattern classifications are discussed. Approaches of EEG signal analysis on the basis of symbolical regression and classification of imaginary movements on the basis of artificial neural networks and support vector machines are considered. It is shown that acceptable for BCI classification accuracy can be attained by using the combination of qualifiers based on neural networks and support vector machines that analyze different feature sets and in case of individual adjustment for users. Peculiarities and perspectives of noninvasive BCI application for post stroke patient rehabilitation, prosthetics and the robotic service for people with the restricted movements are discussed.
