You are on page 1of 16

Scale-Space Wave SCIENTIFIC AND TECHNOLOGICAL OBJECTIVES Stochastic Resonance Synergetics is a generalized description of an intelligent agent, with a memory

and means of action, and its interactions with the environment. It studies coding, control, and propagation of information. This paradigm of computation recognizes an internal state of networked neural system and its interaction with the environment via coupled information propagation. It combines spatial and a dimension of scale in a dynamical system of computation, in the scale-space. The SRS hypothesis aims for building a study discipline of intelligent agents that uses a quantitative analysis of attention, memory and behavioral data. It requires a cross psucho-physics, bio-imaging and computer science, machine learning approaches to achieve:

A proof of the concept in multi-dimensional, information-driven situations: unpredictable and/or evolutions in a systemic or environmental nature; both, a system of intelligent agents and environments may entrain each other at different spatio-temporal and dynamic scales. Diagnosis of disordered states and self-adjustment to those aligned with the environment. Coordination of motion of the self-adjusted intelligent agent systems for different (structural as well as functional) morphodynamic and evolutionary purposes.

The main goal of the project is to further develop and validate scale-space approaches using renormalization techniques that yield hierarchies of robust regimes of synergetics

of agents and environments. The characteristics of these hierarchies provide insights in the (non-)functional and organizational properties of such synergetic networks across spatio-temporal and dynamic scales. These properties in turn can serve as system requirements in purposefully artificial intelligence modeling and engineering. Dimensional interchangeability is explored in memory organization and pathway programming. Decomposition of an ensemble statistics in a set of clusters, and its multiscale organization, is a rich model description with spatio-temporal events for memory organization, as well as, motor program learning and coding. Simulation and experimental techniques are proposed in assessing trade-offs between intelligent agents learning and constrained motor programs coding. The scientific and technological objectives are: 1. The theoretical foundation of stochastic resonance synergetics as a robust and scalable solution to adaptation and learning, coding, perception, attention and movement control of intelligent systems in uncertain and evolving environments.

2. Attention, memory organization and behavioral data-driven study as a result of the motion and coordination planning with the quantitative analysis shown in as a solution to such the theoretical model proposition. 3. Parallel, distributive algorithmic research and proof-of-concept computer implementation of artificial intelligence renormalized synergetics, respectively. Research methodology A multi-scale method of polynomial complexity has been derived for robust multidimensional data analysis [ref. 17.]. This computational method captures a physical model for the generation of the underlying data. Despite its non-linear and dynamical nature, the method yields a model-consistent and concise hierarchical description of the observed multidimensional data. An example of motion through the scale-space is shown in Figure 1. At the scale equilibrium (+=-), the fixed points computation avoids local minima. The information is propagated by the wave equation:

Figure 1: The wave information propagation (3) enables motion towards a global minima. The free energy F and distortion (variance) V are computed for a cluster of data points within a spatial window W. Computation of the spatial vector is coupled by the scale parameter [ref: 17.]. Mass conservation principle, along with the generalized uncertainty and the scale-space wave propagation, leads to a polynomial decomposition of information [ref: 10.]. We propose it within a quantum information theory that lays down a new perspective to networked systems dynamics and computation. We place a special attention in this proposal to the hypothesized questions: how and why information mass-couples in the clusters/quanta of information? At a fixed point, cluster dynamics of data points is kept in a synchrony, and the computed manifolds map its structure in the scale-space domain. Although, a valuable tool in analyzing network statistics, the search for its fixed points, we find it essential, to complement with the model perturbation propagate in the scale-space (3). In particular,

we are interested in applications in network pathways computing and the associated control structure. We, therefore, denote this methodology as a scale-space computing. It is programmable, as an assumed ensemble statistics, but at the same time adaptable and evolving with a particular realization to the environmental signals. Computation is carried out by the scale-space wave information propagation. At the point of a wave collapse, + = - = 0, a quantum of information is split in two. On the other hand, the critical point of dynamical stability, + = - 0, defines the condition of wave resonance, and quantizing the information within a nucleon. Conditions of static stability are defined by the positive definitiveness of the free energy for the quanta of information, and the synergy coupling within -1 +1. A computable set of spatiotemporal events, organized in a graph structure, is proposed as a coding scheme. The equilibrium points in the segmented search across the hierarchy of scales parameterize the node points of the graph structure. The coupling parameter quantify the synergy exchange while traversing the graph in the scale-space computing pathway. An example decomposition is given in Figure 2.

Figure 2: An example graph structure of the scale-space decomposition scheme (From ref. [17]).

We studied complexity of brain wave recordings during a sleep with such a coding scheme, to show expressiveness of quantum information. For an example, it clearly shows the REM phases with the lightest quanta being shut, in figures 4.(a) and (b).

Figure 3: A single nucleon decomposition, at -1 < < +1. Two scale-space waves resonate information within a nucleon. Their positive/negative envelopes are depicted in the color graphs, where green/red belongs to one, and yellow/blue to another wave. (From ref. [1]).

(a)

(b)

Figure 4: Two nucleons decomposition at the limit of the static stability, =1. Short term events, associated here with the higher voltage peaks in signal, will dominantly cluster in the heavier quanta of information (green). The lightest quanta (yellow) cluster unipolar voltage values, negative in the nucleon (a), and positive in (b). (From ref. [1]). To achieve the scientific and technological objectives A-C, we extend and apply scalespace computing to the problem of qualifying and quantifying measures for making operational sustainable synergetics of agents and environment. Well show how the solution to NP hard problems, like traveling salesman problem for motion planning, holds an answer to the quantitative analysis of attention, memory and behavioral datadriven study. Thereto, on the one hand the foundations of synergetics are further explored, in particular for evolving networks of modular intelligent agents and environments. Such networks and environments in contrast with normally studied systems live on a space of order parameter field equations. It is uncertain which (sub-) system is enslaving / entraining which and how as such systems are by nature open. In addition, the renormalization techniques based on our scale-space approach will be tuned to the physics of the synergetics of such open systems. On the other hand, for validation purposes scale-space simulation and experimental techniques are developed for qualifying and quantifying the synergetics. Numerical aspects such as the consistency of the renormalization schemes with the observed synergetics will be

studied with special attention. The proposed method will be made suitable for the implementation in parallel computer architectures. In particular, we study hierarchies of renormalized synergistic learning, coding, and control of systems. We intend to research these in an unified way as a dynamical system.

Figure 5: Hierarchical spatio-temporal events movements coding. (From ref. [16]). The ensemble of joints movement trajectories during a reaching task are shown in Figure 5. A hierarchical coding scheme, with spatio-temporal events, is shown for coding the ensemble statistics of the movement. Note that in Figure 5, Gaussian profile of the ensemble statistics approximates the joints angular velocity trajectories. It gives, therefore, a way of coding the statistics of an ideal movement, for a selected set of joints. Generally, the more complex movements are learned for a given set of joints within the dynamical system. In the tripartite scheme (learning, coding, and control), the movement learning is dominantly temporal problem.

We foresee the movements coding as dominantly spatial problem, suggesting synergistic segmented control, governed by the multi-scale wave dynamics. Although the dynamical scheme does not distinguishes separately the learning, coding, and control part of the scheme, it can be readily assumed that the learning movement structure takes part on much slower time scale, as compared to actually performing the movement in accordance to a pre-coded movement plan. Information theoretical consideration of a scale-space differ from that of the Minkowski space-time continuum in a way that it makes information decoupling in space more explicit, giving the way of the quantum information. Dimensional interchangeability is explored in memory organization and pathway programming. Decomposition of an ensemble statistics in a set of clusters, and its multi-scale organization, is a rich model description with spatio-temporal events for memory organization, planning, as well as, motor program learning and coding. Simulation and experimental techniques are proposed in assessing trade-offs between learning and constrained motor programs coding, for the best model performance of a network of intelligent agents. Progress beyond State of the Art Information theoretical approach to statistical mechanics was introduced in the work of E.T.Jaynes, and the principle of maximum entropy was proposed as an inference procedure. Clustering techniques are applied in many problems (like pattern recognition, learning, source coding, image and signal processing) where a priori knowledge about the distribution of the data is not available. Clustering of data is a widely used tool for analyzing multidimensional data in diverse disciplines such as engineering, physics, biology, and social science. Various approaches to the probabilistic and fuzzy inference in clustering are presented in the literature. The computation of the clustering parameters becomes more effective on the adaptively selected, local windows of computation in our clustering algorithm, as compared to the other techniques. This translates to a network structure segmented by distinct sets of the processing units, with the associated pattern vectors, and dynamic coupling parameters. Distributive modular network of intelligent agents can take advantage of such a parallelization in the computation, on the one end, and in the motion mechanics coding and control, on the other end of the application.

A multi-fractal model formalism was derived in the Thalweg ARC project [ref. 10.], to achieve a harmonic decomposition of image sequences, and its singularity analysis. The partition function describes the probabilistic model of data clusters and is analyzed as a multi-fractal measure in the method. Singularity analysis of the computed maps of clustering vectors is used to explain the computational means of decomposing a data base into different singular sets. We showed that the propagation of information in data sequences is governed by the scale-space wave equation, therefore enabling us to treat singular frequencies of data clusters in an unified way, both in space and in time. Contextual information of the spatial coherency of data is used in the segmentation process in the hierarchical scale computation of feature vectors. The spatial segmentation is performed while using the Greens function, parameterized with the scale parameter, as the integration function in the segmentation process. The scale information is evaluated by conjoining the two parameters: the scale parameter of the signal distortion, and the spatial scale parameter r. A larger extent of spatial integration of the data information is used on a larger scale, while it becomes effectively more local in space as we decrease the scale of computation. Our proposed method is deterministic and well posed, along with its polynomial complexity. Distinct singular features are segmented on a certain scale and the most singular features obey the Laplacian system regularity constraints, in the hierarchy of the scale computation. Accordingly, the reconstruction formula is derived based on the Laplacian system of coupled manifolds spanning the content of information within clusters. This gives us an effective way of compressing and progressive coding and transmission of information in data sequences.. Let us provide here examples of the applicability of our method to a basic intelligence systems problems. Cluster synchronization:A test image sequence of an expanding ball is used as an example pattern for 2-dimensional signal decomposition. The statistical maps of the clusterized 3 image segments are shown in Figure 6. Differential operators: divergence, rotor, and the combination of the two are used for clustering data. The resulting membership distributions of data points are shown on Figures 6.(a-c) for these operators, respectively.

(a)

(b)

(c)

Figure 6: Clustering of the expanding ball images sequence: 3 clusters statistical maps projections along 2D principal components with: (a) divergence, (b) rotor, and (c) combined divergence and rotor point operators. (From the pub. [7]). From a perspective of the information theory, the synchronous dynamics of a system of 3 clusters is related to its 3D scaling properties. We intend to explore and extend this clusters synchronization technique further in a network of intelligent agents analysis, and its control in interaction with an environment. Prediction and Action Modeling: We are interested in exploring dimensional interchangeability (conditionally: space-time) in the search, adaptive memory organization, retrieval of information. From a perspective of the information theory and the synchronous dynamics of a system of 3 clusters, as shown in the previous example, we have worked out a method of keeping in sync system of 3 data clusters. This translates in a way of organizing: 1. event, 2. short-term, and 3. long-term memory of data. Further structuring is possible, as well, through our method of multi-scale decomposition of data. In a motion planning, it structures functionally in windows of attentions nucleons of information, to achieve a predefined goal. It is an approach of polynomial complexity, and with a set of local operators that, in our view, makes it ideal for the implementation on parallel computer architectures. With our approach the synergetics is treated in scale as well as in space-time. This results in clustering spatio-temporal features of synergetics into distinct singular sets.

The multi-scale decomposition of data therewith enables besides perception on different scales, also progressive coding, and control. This forms a novel approach to autonomous pathway programming for distributive modular intelligent systems. Timeliness and relevance of the project A multi-scale method of polynomial complexity is proposed for segmented fixed point search and the manifold computing of multidimensional data. The method is based on a physical model of signal distortion. Although the non-linear and of a dynamic nature, the method aims for the simplest decomposition explanation and the computation of multidimensional data. Its near-linear characteristics are given an extended theoretical considerations. Dimensional interchangeability and pathway analysis make this method a tool for data mining, coding and control of multidimensional data in a lowdimensional scaled domain. It is also suitable for the implementation in parallel computer architectures. Next generation networked intelligent systems and the new application domains are subject of our scientific research and implementation design. Multi-scale renormalization system behavior, built in the operating system of the network of intelligent agents, brings state of the art research in the artificial intelligence, further. This approach, in its theoretical foundation, links a dynamical systems behavior to a more complex computational modeling of natural cognition in symbiotic organisms. Such a multi-disciplinary scientific bilingualism, we believe, develops an intuitive evaluation framework for assessing the renormalized synergistic performance. RESEARCH RESULTS We find a scale-space computing paradigm applicable with the scale-space wave propagation of the uncertainty of information, as used in Dr. Milan Jovovic reports. We find it important to analyze dimensionality issues, as well, along with the computer implementation. As a visiting scholar at the Indiana University, Dr. Milan Jovovic has worked on multi-core processors research and parallel algorithms. The emphasis was on applications in chemo/bio-informatics and data mining. He has explored covariant differentiability in approaching the scale-space computing, knowledge diagrams and multi-scale numerical schemes for the computation of the high-dimensional data sets. An implementation within the 3D statistical maps is analyzed in his report.

On the posdoc stage at the INRIA, Dr. Milan Jovovic has worked on harmonic signal decomposition of the IR satellite images. Signal scale invariance is used in detecting and tracking the rain patterns in convective clouds. He has proposed, also, learning of limb movements, in humans, as a sequence of synergistic joint motions, at the University of Belgrade. Back in the 1990s, Dr. Milan Jovovic has been involved in the study of neurophysiology. He has proposed segmentation of visual textures dynamically, by a recurrent scheme. He has worked, also, on topics such as visual attention, pathways, and the sensory-motor integration. We intend to extend this research in the swarming intelligence and robotics analysis and control. We propose multi-scale synergistic modeling and computation along with the associated renormalization numerical scheme, and computer implementation. Publications: 1. Jovovic, M., [2011], Brain wave synergies, analysis and coding 2. Jovovic, M., [2010], Foundation of re-normalized synergetics issues of computability and complexity, Poster presentation, New England Complex Systems Institute. 3. Jovovic, M., A. Salden, [2009], Foundation of Renormalized Synergetics applied to Robotics, Marie Curie Fellowship Application. 4. Jovovic, M., [2009], Multidimensional Data Scaling, and the Scale-Space Computing Approach, Poster presentation, IDSIA Lab, Switzerland. 5. Jovovic, M., [2009], Multidimensional data scaling: analysis, coding and control, Poster presentation, Microsoft Research, Cambridge, UK. 6. Jovovic, M., [2008], Where dimensions come from? dynamical approach to scaling, Poster presentation, Belgrade University, Serbia. 7. Jovovic, M., G. Fox, [2007], Multi-dimensional data scaling dynamical cascade approach, Technical Report Indiana University.

8. Jovovic, M., [2007], Scale-space computing: analysis, coding, and control, Poster presentation, Indiana University, USA. 9. Jovovic, M., [2006], Texture segmentation by 2-layers recurrent dynamics, XXVIIIth International Symposium in COMPUTATIONAL NEUROSCIENCE, Montreal. 10. Jovovic, M., H. Yahia, I. Herlin, [2003], Hierarchical scale decomposition of images singular feature analysis, Technical Report INRIA, AIR Lab. 11. Jovovic, M., [2002], Hierarchical scale quantization and coding of motion information in image sequences (in English), Informacione Tehnologije VI, Zabljak, Yugoslavia. 12. Jovovic, M., [2002], Space-Color Quantization of Multi-Spectral Images, Poster presentation, EPFL, Switzerland. 13. Jovovic, M., [2001], Space-Color Quantization of Multispectral Images in Hierarchy of Scales, Int. Conf. on Image Processing, Thessaloniki, Greece, pp. 914-917. 14. Jovovic, M., [2001], A multiscale space-color quantization of images, Informacione Tehnologije VI, Zabljak, Yugoslavia. 15. Jovovic, M., et al., [2000], Telemedicine: shared information in health systems. Informacione Tehnologije V, Zabljak, Yugoslavia. 16. Jovovic, M., et al. [1999], Automatic synthesis of synergies for control of reaching hierarchical clustering. Medical Engineering and Physics 21/5:325-337. 17. Jovovic, M., [1999], A multiscale processing of data streams (in English). Informacione Tehnologije IV, Zabljak, Yugoslavia. 18. Jovovic, M., [1999], Automatic control of reaching: nonparametric learning. Ph.D. thesis, University of Belgrade. 19. Jovovic, M., D. Popovic, [1998], Hierarchical clustering for determining the reaching synergies. Informacione Tehnologije III, Zabljak, Yugoslavia.

20. Jovovic, M., [1996], Image segmentation for feature selection from motion and photometric information by clustering. SPIE Symposium on Visual Information Processing V, Orlando. 21. Jovovic, M., [1994], Texture Discrimination by Adaptive Filtering. 17th. European Conference on Visual Perception, Eindhoven. 22. Jovovic, M., [1994], A Markov random fields model for describing unhomogeneous textures: generalized random stereograms. IEEE Workshop Proceedings on Visualization and Machine Vision, and IEEE Workshop Proceedings on Biomedical Image Analysis, Seattle. 23. Jovovic, M., [1994], Shape and motion recovery from optical images in lightattenuating media. Technical Report, University of Miami. 24. Jovovic, M., [1990], Attentional Representation and Learning of Hand-written Characters. Technical Report, Caltech. Independent thinking and leadership qualities Dr. Milan Jovovic, coordinator, have participated in various projects, mostly described as multidisciplinary with a wide team of specialists in different disciplines. He has acquired an impressive experience and a broad theoretical knowledge. His contributions have been described as an innovative researcher, with independent thinking that results in his recognizable research profile (please see publication list). We believe his profile fits well to our joint approach in bringing about the foundation of renormalized synergetics applied to robotics, with this project. He will be taking a leading role in coordinating theoretical development, software implementation, and publication of results. Match between the fellows profile and project Dr. Milan Jovovics scientific endeavor has been dominated by the idea of an universal computational paradigm governing information processing by neural systems as within physical systems. He has started working on it when moved from the Prof. Geoffrey Fox to Prof. David Van Essen lab, at the Caltech. He has considered the ways to combine

spatial and dimensions of scale in diffusing the information between two levels of potential energies. His intentions were to generalize description of an intelligent agent, with a memory and means of action, and its interactions with the environment. This paradigm of coupled information propagation gives a way of motion through the scalespace, and therefore computation. The study of coding, control, and transfer of information Dr. Milan Jovovic finds as his entrance points to the fields of theoretical neuroscience, as well as, artificial intelligence. He is very much interested in a research that would generalize it within a string-type logic formalism. Potential for reaching a position of professional maturity Dr. Milan Jovovic has worked in information engineering (for about 20 years), from different fields of study and perspective. Dr. Milan Jovovic has acquired a broad theoretical knowledge in the fields. He has reached a significant level of mathematical elaboration in his work, to express his contributions. We expect that could qualify his research for further funding. Potential to acquire new knowledge Dr. Milan Jovovic loves challenges in science and research. He has tried to make his living in industry, even to find his ways in the turmoil in the Balkans. We believe, though, that the environment and a job in the research and academia suits him the best. We would be happy if he is given the chance to prove his abilities. In mathematics, Dr. Milan Jovovic prefers descriptions through canonical network forms knowledge diagrams and the coupled systems dynamics. He would love, also, to understand better cognitive systems, especially language processing. He is interested in improvements in robotics, control, and design. The applications in biomedical engineering (genome analysis, molecule modeling and synthesis) and social sciences are exciting possible areas of research to him, as well. Contribution to career development or re-establishment where relevant Dr. Milan Jovovic, have shown consistency throughout his career in developing dynamical state modeling of computational neural systems, as well as, complex systems analysis originating in computational physics. We expect the opportunity of this project

participation will develop his career further, to a more mature state that will reestablish his position of a senior researcher in these fields. Contribution to European excellence and European competitiveness The research on renormalized synergetics will put Europe at the forefront of physicsbased robotics. The results will allow science, industry and society gaining valuable information about the hierarchies of renormalized sustainable synergetics of robots and environments. In particular the robotic performance under different environmental pressures of a dynamic or evolutionary nature and across various levels of scale will become assessable. This will allow Europe to establish a competitive edge in building up robots in interaction and co-evolution with environments from the nano up to the macro scale. Benefit of the mobility to the European Research Area Scale-space or multi-scale physics computing is a research area quite underdeveloped either within Europe or the rest of the world. In order to achieve major scientific and technological breakthroughs in particular in robotics the need for advancement of scalespace computing becomes evermore important. Classical system engineering approaches based on e.g. evolutionary computing, not taking any empirical synergetics model into account through e.g. data assimilation techniques, will show to fall short. Dr. Milan Jovovic is one of the exceptional candidates that can bring the state of the art in this domain further, and will succeed by being allowed to collaborate with the very few experts in this domain and outstanding groups within robotics.

You might also like