H. Ritter, T. Martinetz, and K. Schulten, Neural Computation and Self-Organizing Maps: An Introduction, Addison Wesley, 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....fashion. The mappings make topological neighborhood relationship geometrically explicit in low dimensional feature map. This makes them interesting for applications in various areas ranging from simulations used for the purpose of understanding and modeling of computational maps in the brain [3] to subsystems for engineering applications such as cluster analysis [4] motor control [5] speech recognition [6] vector quantization [7] adaptive equalization [8] and combinational optimization [9] Kohonen et al. 10] provided partial reviews. Manuscript received March 30, 1998; revised ....

H. J. Ritter, T. Martinetz, and K. J. Schulten, Neural Computation and Self-Organizing Maps: An Introduction. Reading, MA: Addison -Wesley, 1992.

....towards the centre of the view or away. For example, Berthouze et al. 27] smooth pursuit controller, based on Feedback Error Learning (FEL) 28] did not use velocity of the target. Also, current artificial ocular motor map techniques for saccadic motion have not considered target velocity [29 32]. Researchers at the LiraLab developed an integrated system, combining various biologically inspired eye and head movement mechanisms [33] Target velocity was not considered, except when compensating for induced pan movements [34] Shibata and Schaal s [18] controller included velocity ....

H. Ritter, T. Martinetz, and K. Schulten, Neural Computation and Self-Organizing Maps: An Introduction, Addison Wesley, 1992.

....preservation and 3. probability density matching and magnification. Convergence and ordering: For continuous inputs Ritter and colleagues investigated the stationary state and convergence properties after ordering under the assumption of a continuous index set A, the results are summarized in [109]. Erwin et al. have proved that it is impossible to associate a global potential function to SOM for continuous inputs and studied the role of the neighborhood function [45, 46] Thereby, the learning is taken as Markov process [66] An intuitive straightforward definition of a potential function ....

....Violations of topology preservation do not only arise because of convergence problems: if the lattice dimension DA differs from the effective data dimension Def t E)v topological mismatches occur. The respective theory of meta and instable states is initially based on Fokker Planck approaches [109]. Further studies also use the Ginzburg Landau theory to describe the phenomena in more detail [36] The highdimensional analysis was pointed out in [11] using phase diagrams. To overcome the topological mismatch problem growing variants of SOM have been developed [12] or input pruning was tried ....

[Article contains additional citation context not shown here]

H. Ritter, T Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley, Reading, MA, 1992.

....distribution of centroids and the initial distribution fix) During the learning phase, the Kohonen algorithm uses neighbourhoods of centroids taken into account during the adaptation; the size v of these neighbourhoods is usually decrased during learning. For v fixed, the algorithm is equivalent [11] to the minimisation of a generalised distortion function: v (f, dP) x yi f(x)dx (3) where F(i) is the set of indexes in the neighbourhood of h including i. This generalised distortion function can also be estimated through a finite set of samples xt, x2, Cv, similarly to (2) 212 ....

Ritter H., Martinetz and Shulten K., Neural Computation and Self-Organizing Maps: an Introduction, Addison-Wesley, Reading, 1992.

....robot must gain knowledge of its motion control, otherwise known as kinematics, in a given environment. This relates to sensorimotor coordination . How then can the mobile robot be acquainted with its own kinematics Our research rooted off by examining the control architectures of robot arms [5, 8, 9, 14, 15, 36, 37] to identify the underlying concepts behind automatic end effector positioning; these concepts may be applicable to our cause. In doing so, we have to observe that the direct translation of learning and sensorimotor control techniques from robot arms to mobile robots is not possible due to several ....

....Many unsupervised learning algorithms have been dedicated to the application of sensorimotor coordination on robot manipulators and mobile robots. They include evolutionary optimization or genetic algorithms [33] rule based algorithms [34] fuzzy logic [35] artificial neural networks [5, 6, 7, 8, 9, 14, 15, 30, 36, 37, 48, 49, 50, 60] and reinforcement learning [38] The choice of the learning algorithm must take into consideration all the objectives stated in Section 1.2. In particular, artificial neural network fits our purpose extremely well; it replicates the cerebellum of a mammalian brain, which performs sensorimotor ....

[Article contains additional citation context not shown here]

Ritter H., Martinetz T. and Schulten K., Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley, Reading, MA, 1992.

.... algorithm can be considered as a generalization of both the GL algorithm and the HS algorithm (with local averaging) The relation of SOM and vector quantization 38 is a widely known fact (see, e.g. Koh97] whereas the similarities between SOM and principal curves were pointed out recently by [RMS92] and [MC95] The SOM algorithm is usually formulated as a stochastic learning algorithm. To emphasize its similarities to the HS and GL algorithms, we present it here as a batch method as it was first formulated by Luttrell [Lut90] In its original form, the SOM is a nearest neighbor vector ....

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley, Reading, Massachusetts, 1992.

....and nal (target) positions. The robot receives sensory information from the workspace and autonomously constructs some kind of inverse mapping. Typical examples of this approach are the works of Grossberg Kuperstein [13] Kuperstein Rubinstein [14] Martinez et al. 15] and Ritter et al. [16]. The three rst works describe a self organizing model for visuomotor coordination of a robot arm. This model learns to control a 5 degree of freedom (DOF) robot arm to reach cylindrical objects. The authors use a set of one dimensional topographic maps that represent the location of the target ....

....each actuator, they can approximate only a restricted class of control laws. The work of Bullock Grossberg [17] extends Kuperstein s model by including muscle dynamics, initial conditions, muscle contraction rates, and feedback signals from muscle sensors. Martinez et al. 15] and Ritter et al. [16] presented an approach to diminish the drawbacks of the Kuperstein model, by using 3D variant of the Kohonen SOM. In this approach the ordering and resolution of the map evolve during learning (by updating a layer of input weights) thus determining the distribution of the neural units over the ....

H. Ritter, T. Martinetz, and K. Schulten. Neural computation and self-organizing maps: An introduction. Addison-Wesley, Reading, MA, 1992.

.... learning strategy that makes use of multiple corrective saccades to adaptively form a retinotopic motor map similar in spirit to the one known to exist in the deep layers of the primate superior colliculus [13] Our approach differs from previous strategies for learning motor maps (for instance, [12]) in that we use the visual modality to actively supply the necessary reinforcement signal required during the motor learning step (Section 3.2) 2 The Multiscale Spatial Filter Representation In the active vision framework, vision is seen as subserving a larger context of the encompassing ....

.... before a goal directed saccade corre sponds to the brightest spot (most likely match) in the distance image (Figure 1 (c) for example) The formation of sensory map can be achieved using Kohonen s well known stochastic learning algorithm by using a Gaussian input density function as described in [12]. Our primary interest lies not in the formation of the sensory map but in the devel opment of a learning algorithm that assigns appropriate motor vectors to each location in the corresponding retinotopically organized motor map. In particular, our algorithm employs a visual reinforcement signal ....

Helge Ritter, Thomas Martinetz, and Klaus Schulten. Neural Computation and Self-Organizing Maps: An Introduction. Reading, MA: Addison-Wesley, 1992.

....system does not currently use these additional methods, although we are considering such combinations. The use of a topological ordering of neurons for selforganization has been used by Kohonen [8] This has been extended to include output learning for control by Ritter, Martinetz, and Schulten [13]. These ideas were brought together with the use of the eligibility trace by Hougen [4] and the present paper is believed to be the first to explicitly consider the trade off between a completely learned partitioning and one assigned by the system designer. ....

H. Ritter, T. Martinetz, and K. Schulten. Neural computation and self-organizing maps: an introduction. Addison-Wesley, Reading, MA, 1992.

....parametros de forma a continuarem operando com efic acia sem preju izo de nenhuma de suas caracter isticas. Num ambiente perfeitamente conhecido e controlado, diversas t ecnicas, tais como programac ao lead by nose , c alculos de cinem atica e dinamica inversas [8] treinamento de redes neurais [9, 14, 16] e uma variada gama de t ecnicas de inteligencia artificial [12] podem ser utilizadas. Mas todas estas t ecnicas apresentam s erias desvantagens quando utilizadas para o controle de um sistema inserido num ambiente aberto [5] Devido a complexidade e n ao linearidade de um sistema de controle ....

....possuir alguma tolerancia a ru ido de informac ao. Falhas eventuais na aquisic ao de informac ao n ao devem interferir significativamente na tarefa em execuc ao. Redes neurais, al em de apresentarem caracter isticas de tolerancia a ru ido, s ao capazes de resolver problemas altamente n aolineares [7, 9, 15], como e o caso de sistemas de controle rob otico. Os sistemas adaptativos, em geral, por forma a resolverem problemas com caracter isticas de n aolinearidade, necessitam de uma quantidade muito grande de parametros, cujos valores devem ser adequadamente atribu idos, permitindo que estes ....

[Article contains additional citation context not shown here]

Ritter, H.; Martinetz, T. e Schulten, K. Neural Computation and Self-Organizing Maps -- An Introduction. Addison-Wesley Publishing Company, 1992.

....most of the time, no longer suitable. The problem of positioning a mechanical arm in the space must be the first one to be tackled. In a controlled, perfectly known environment, several techniques, such as lead by nose programming, inverse kinematics calculations [7] trained neural networks [8, 13, 15] and a large gamma of classical artificial intelligence methods [11] can be used. But those techniques present some serious drawbacks when used within an unknown environment [5] The majority of solutions to position a robot arm try to map a spatial position and orientation to a joint ....

....the previous and current values of those variables. The combination of the previous two constraints make the convergence slower, but they are necessary to prevent possible collisions or undesirable oscillations. ffl Relaxation constraint Humans are known for using lazy arm configurations [8], that is, they tend to use configurations that are not completely extended or fully contracted. If it is possible, we set our joints with middle range values. The result of such a procedure is a mechanical arm in a relaxed position from a human point of view. This constraint is also important ....

Ritter, Helge; Martinetz, Thomas e Schulten, Klaus, Neural Computation and Self-Organizing Maps -- An Introduction. Addison-Wesley Publishing Company, 1992.

....or mutual inhibition) all units will merge together on the expectation value of the input, E(x) Let us observe how some algorithms adapting more than one unit per step unlink adaptation function G(i; x) from the euclidean distance: 3.3. 1 Kohonen algorithm In Kohonen s self organizing maps [8, 9, 10, 15, 14], once the euclidean winner has been found, the adaptation factor is computed in function of the lateral distance d(i; w) along the grid, where w is the winner index: G(i; x) exp( Gammad 2 (i; w) 2 ) 7) gaussian neighborhood) where is a neighborhood radius decreasing with the time, or: ....

H. Ritter, T. Martinetz, and K. Schulten. Neural computation and self-organizing maps: an introduction. Addison-Wesley Publishing Company, 1992.

.... neural network historically inspired by sensory maps found in biology ( 17, 11, 12] They are well known for their ability to provide a data mapping where both sample number and dimensionality are reduced, and can be viewed as a non linear extension of Principal Component Analysis (PCA) [2, 15, 14]. However, SOM have a major drawback: the mapping is done toward a grid of neurons whose shape is a priori fixed and may not comply with the one of the data submanifold, leading to confuse mapping. The grid is generally rectangular or hexagonal, and for most applications (especially with ....

H. Ritter, T. Martinetz, and K. Schulten. Neural computation and self-organizing maps: an introduction. AddisonWesley Publishing Company, 1992.

....version [89] the neighborhood function h r was equal to 1 for a certain neighborhood of i, and 0 elsewhere. The neighborhood and the gain fl should slowly decrease in time) The convergence and the mathematical properties of this algorithm have been considered by several authors, e.g. 93] and [139]. The role of the SOM in feature extraction is to construct optimal codewords in abstract feature spaces. Individual feature values can then be replaced by these codes, which results in data compression. Furthermore, hierarchical systems can be built in which the outputs from the maps are again ....

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley Publishing Company, Reading, Massachusetts, 1992. 57

....distribution of centroids and the initial distribution f(x) During the learning phase, the Kohonen algorithm uses neighbourhoods of centroids taken into account during the adaptation; the size v of these neighbourhoods is usually decrased during learning. For v fixed, the algorithm is equivalent [11] to the minimisation of a generalised distortion function: x n v i C i n f x y f x dx k k V , F = 2 1 U (3) where V(i) is the set of indexes in the neighbourhood of I, including i. This generalised distortion function can also be estimated through a finite set of ....

Ritter H., Martinetz and Shulten K., Neural Computation and Self-Organizing Maps : an Introduction, Addison-Wesley, Reading, 1992.

.... Neural network quantization of 3D biological data Self organizing feature maps organize the connectivity of neurons in a cortical layer to optimize the spatial distribution of the neural responses to a training sequence of input signals (Willshaw von der Malsburg, 1976; Kohonen, 1982; Ritter et al. 1992). The purpose of the optimization is to convert the similarity of signals into proximity relationships among the neurons in the cortical layer. Self organizing neural networks are typically applied to solve important information processing tasks in articial intelligence, including the formation of ....

....of signals into proximity relationships among the neurons in the cortical layer. Self organizing neural networks are typically applied to solve important information processing tasks in articial intelligence, including the formation of topographic sensory maps and visuomotor control of robots (Ritter et al. 1992; Zeller et al. 1997) A consequence of the map formation is a quantization that approximates the probability density function p(v) of input signals v 2 D , using a nite number of neural pointers w i 2 D , i 1, N. Here, we train neural networks on 3D biological data sets, using the ....

[Article contains additional citation context not shown here]

Ritter, H., Martinetz, T. & Schulten, K. (1992). Neural Computation and Self-organizing Maps: An Introduction, Addison-Wesley, New York.

No context found.

H. Ritter, T. Martinetz, and K. Schulten, Neural Computation and Self-Organizing Maps: An Introduction, Addison Wesley, 1992.

No context found.

Ritter, H. J., Martinez, T., and Schulten, K. J. (1992). Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley, Reading, Massachusetts.

No context found.

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps - An Introduction. Addison-Wesley publishing company, 1992.

No context found.

H. Ritter, T. Martinetz, and K. Schulten, Neural Computation and Self Organizing Maps: An Introduction Addison-Wesley: Reading, MA, 1992.

No context found.

H. Ritter, T. Martinetz, and K. Schulten. Neural computation and self-organizing maps: an introduction. Addison-Wesley, Reading, MA, 1992.

No context found.

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley, Reading, MA, 1992.

No context found.

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps: An Introduction, Addison-Wesley, 1992.

No context found.

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and Self-Organizing Maps: An Introduction. Addison-Wesley, Reading, MA, 1992.

No context found.

H. Ritter, T. Martinetz, and K. Schulten. Neural Computation and SelfOrganizing Maps: An Introduction, Addison-Wesley, 1992.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC