Project of the Scientific Grant Agency VEGA 2/0019/10

Duration of the project: 01/2010 - 12/2012

Principal investigator: prof. RNDr. Ing. Ivan Bajla, PhD

Project partners: Faculty of Mathematics, Physics and Informatics of the Comenius University Bratislava (FMFI UK) and Institute of Measurement Science of the Slovak Academy of Sciences (IMS SAS)

Project summary  

The project is oriented towards exploration and applications of novel prediction methods of nonlinear dynamic systems, biologically inspired hierarchical network models (HTM) with inherent temporal component, artificial neuronal networks and advanced mathematical statistics for analysis of time series (mainly biosignals), and object recognition in visual scenes. Its key goals are: to explore dynamics of complex time series from various experimental domains and to propose original prediction algorithms, to explore the methods of computer vision based on the new prediction-memory network model of HTM, to apply the developed algorithms to tasks of object recognition in complex visual scenes comprising moving and occluded objects and to prediction of one- and multi-dimensional biological time series. 

State-of-the-art of the project in January 2012  

A special attention was devoted to problems of predictability of synthesized  and real time series. The nonlinear prediction, based on the reconstruction of dynamical properties of the system being studied, proved to be succesful. An optimum reconstruction requires to determine the smallest space dimension that is sufficient for the construction of an equivalent dynamical model. In this relation we analysed a recently introduced notion of „observability“, and proposed a modified method. This method determines a sufficient dimension of inclusion from the decrease of the number of the so-called false neighbors in the process of reconstruction dimension increase. We tested a number of conventional and advanced measures for the analysis and classification of sleep signals (EEG, EKG, EOG a EMG). The best classification was achieved using a combination of the EEG power in delta, alpha and beta bands. The outstanding classification capability was also manifested by some novel characteristics, like fractal dimension and exponent of power spectrum decrease.

The further exploration of the HTM model applied to the classification problem of geometric objects lead us to the problems of optimal codebook generation and to the development of a more efficient temporal pooler operations. We replaced the initial approach of Numenta, the so-called smooth explorer, by a novel „random walk“ approach that is based on the Metropolis-Hastings algorithm of the generation of a Markov sequence of samples from arbitrary probability distribution. By introducing a suitable likelihood function into this algorithm, we significantly decreased the number of learning steps of the given network. We have also re-formulated the initial algorithm of inference in HTM network defined by Gaussian function. Thereby a reduction of controlling parameters has been reached. The core contribution of this research is represented by the development of a pair-wise explorer (temporal pooler) which yields the temporal groups more efficiently than previously. The computer experiments with one-level HTM network using the modified operations showed that the pair-wise explorer yields faster convergence to theoretical maximum of classification rate on the given set of geometrical objects.

Publications