Neural networks: the official journal of the International Neural Network Society

Published by Elsevier
Online ISSN: 0893-6080
Publications
Article
The chaotic neural network constructed with chaotic neurons exhibits rich dynamic behaviour with a nonperiodic associative memory. In the chaotic neural network, however, it is difficult to distinguish the stored patterns in the output patterns because ...
 
Article
A fundamental capacity of the perceptual systems and the brain in general is to deal with the novel and the unexpected. In vision, we can effortlessly recognize a familiar object under novel viewing conditions, or recognize a new object as a member of a familiar class, such as a house, a face, or a car. This ability to generalize and deal efficiently with novel stimuli has long been considered a challenging example of brain-like computation that proved extremely difficult to replicate in artificial systems. In this paper we present an approach to generalization and invariant recognition. We focus our discussion on the problem of invariance to position in the visual field, but also sketch how similar principles could apply to other domains.The approach is based on the use of a large repertoire of partial generalizations that are built upon past experience. In the case of shift invariance, visual patterns are described as the conjunction of multiple overlapping image fragments. The invariance to the more primitive fragments is built into the system by past experience. Shift invariance of complex shapes is obtained from the invariance of their constituent fragments. We study by simulations aspects of this shift invariance method and then consider its extensions to invariant perception and classification by brain-like structures.
 
Article
This paper presents an on-line unsupervised learning mechanism for unlabeled data that are polluted by noise. Using a similarity threshold-based and a local error-based insertion criterion, the system is able to grow incrementally and to accommodate input patterns of on-line non-stationary data distribution. A definition of a utility parameter, the error-radius, allows this system to learn the number of nodes needed to solve a task. The use of a new technique for removing nodes in low probability density regions can separate clusters with low-density overlaps and dynamically eliminate noise in the input data. The design of two-layer neural network enables this system to represent the topological structure of unsupervised on-line data, report the reasonable number of clusters, and give typical prototype patterns of every cluster without prior conditions such as a suitable number of nodes or a good initial codebook.
 
Article
Humans demonstrate a remarkable ability to generate accurate and appropriate motor behavior under many different and often uncertain environmental conditions. In this paper, we propose a modular approach to such motor learning and control. We review the behavioral evidence and benefits of modularity, and propose a new architecture based on multiple pairs of inverse (controller) and forward (predictor) models. Within each pair, the inverse and forward models are tightly coupled both during their acquisition, through motor learning, and use, during which the forward models determine the contribution of each inverse model's output to the final motor command. This architecture can simultaneously learn the multiple inverse models necessary for control as well as how to select the inverse models appropriate for a given environment. Finally, we describe specific predictions of the model, which can be tested experimentally.
 
Article
A recurrent neural network can possess multiple stable states, a property that many brain theories have implicated in learning and memory. There is good evidence for such multistability in the brainstem neural network that controls eye position. Because the stable states are arranged in a continuous dynamical attractor, the network can store a memory of eye position with analog neural encoding. Continuous attractors in model networks depend on precisely tuned positive feedback, and their robust maintenance requires mechanisms of synaptic plasticity. These ideas may have wider scope than just the oculomotor system. More generally, the internal models postulated by theories of biological motor control may be recurrent networks with continuous attractors.
 
Article
In this paper, the problems of stability of delayed neural networks are investigated, including the stability of discrete and distributed delayed neural networks. Under the generalization of dropping the Lipschitzian hypotheses for output functions, some stability criteria are obtained by using the Liapunov functional method. We do not assume the symmetry of the connection matrix and we establish that the system admits a unique equilibrium point in which the output functions do not satisfy the Lipschitz conditions and do not require them to be differential or strictly monotonously increasing. These criteria can be used to analyze the dynamics of biological neural systems or to design globally stable artificial neural networks.
 
Article
We suggest that any brain-like (artificial neural network based) learning system will need a sleep-like mechanism for consolidating newly learned information if it wishes to cope with the sequential/ongoing learning of significantly new information. We summarise and explore two possible candidates for a computational account of this consolidation process in Hopfield type networks. The "pseudorehearsal" method is based on the relearning of randomly selected attractors in the network as the new information is added from some second system. This process is supposed to reinforce old information within the network and protect it from the disruption caused by learning new inputs. The "unlearning" method is based on the unlearning of randomly selected attractors in the network after new information has already been learned. This process is supposed to locate and remove the unwanted associations between information that obscure the learned inputs. We suggest that as a computational model of sleep consolidation, the pseudorehearsal approach is better supported by the psychological, evolutionary, and neurophysiological data (in particular accounting for the role of the hippocampus in consolidation).
 
Article
In learning goal-directed behaviors, an agent has to consider not only the reward given at each state but also the consequences of dynamic state transitions associated with action selection. To understand brain mechanisms for action learning under predictable and unpredictable environmental dynamics, we measured brain activities by functional magnetic resonance imaging (fMRI) during a Markov decision task with predictable and unpredictable state transitions. Whereas the striatum and orbitofrontal cortex (OFC) were significantly activated both under predictable and unpredictable state transition rules, the dorsolateral prefrontal cortex (DLPFC) was more strongly activated under predictable than under unpredictable state transition rules. We then modelled subjects' choice behaviours using a reinforcement learning model and a Bayesian estimation framework and found that the subjects took larger temporal discount factors under predictable state transition rules. Model-based analysis of fMRI data revealed different engagement of striatum in reward prediction under different state transition dynamics. The ventral striatum was involved in reward prediction under both unpredictable and predictable state transition rules, although the dorsal striatum was dominantly involved in reward prediction under predictable rules. These results suggest different learning systems in the cortico-striatum loops depending on the dynamics of the environment: the OFC-ventral striatum loop is involved in action learning based on the present state, while the DLPFC-dorsal striatum loop is involved in action learning based on predictable future states.
 
Article
This paper examines the multilayer perceptron (MLP) network from a hidden layer decision region perspective and derives the output layer and hidden layer weight constraints that the network must satisfy in performing a general classification task. This provides a foundation for direct knowledge discovery from the MLP, using a new method published by the author, which finds the key inputs that the MLP uses to classify an input case. The knowledge that the MLP network learns from the training examples is represented as ranked data relationships and induced rules, which can be used to validate the MLP network. The bounds of the network knowledge are established in the n-dimensional input space and a measure of the limit of the MLP network knowledge is proposed. An algorithm is presented for the calculation of the maximum number of hidden layer decision regions in the MLP input space.
 
Article
A fundamental problem in neural network research, as well as in many other disciplines, is finding a suitable representation of multivariate data, i.e. random vectors. For reasons of computational and conceptual simplicity, the representation is often sought as a linear transformation of the original data. In other words, each component of the representation is a linear combination of the original variables. Well-known linear transformation methods include principal component analysis, factor analysis, and projection pursuit. Independent component analysis (ICA) is a recently developed method in which the goal is to find a linear representation of non-Gaussian data so that the components are statistically independent, or as independent as possible. Such a representation seems to capture the essential structure of the data in many applications, including feature extraction and signal separation. In this paper, we present the basic theory and applications of ICA, and our recent work on the subject.
 
Article
How we decide whether a course of action is worth undertaking is largely unknown. Recently, neuroscientists have been turning to ecological approaches to address this issue, examining how animals evaluate the costs and benefits of different options. We present here evidence from rodents and monkeys that demonstrate the degree to which they take into account work and energetic requirements when deciding what responses to make. These calculations appear to be critically mediated by the anterior cingulate cortex (ACC) and mesolimbic dopamine (DA) pathways, with damage to either causing a bias towards options that are easily obtained but yield relatively smaller reward rather than alternatives that require more work but result in greater reward. The evaluation of such decisions appears to be carried out in systems independent of those involved in delay-discounting. We suggest that top-down signals from ACC to nucleus accumbens (NAc) and/or midbrain DA cells may be vital for overcoming effort-related response costs.
 
Article
This paper presents a complex-valued version of the back-propagation algorithm (called 'Complex-BP'), which can be applied to multi-layered neural networks whose weights, threshold values, input and output signals are all complex numbers. Some inherent properties of this new algorithm are studied. The results may be summarized as follows. The updating rule of the Complex-BP is such that the probability for a "standstill in learning" is reduced. The average convergence speed is superior to that of the real-valued back-propagation, whereas the generalization performance remains unchanged. In addition, the number of weights and thresholds needed is only about the half of real-valued back-propagation, where a complex-valued parameter z=x+iy (where i=-1) is counted as two because it consists of a real part x and an imaginary part y. The Complex-BP can transform geometric figures, e.g. rotation, similarity transformation and parallel displacement of straight lines, circles, etc., whereas the real-valued back-propagation cannot. Mathematical analysis indicates that a Complex-BP network which has learned a transformation, has the ability to generalize that transformation with an error which is represented by the sine. It is interesting that the above characteristics appear only by extending neural networks to complex numbers.
 
Current conveyors: (a) with comparator, (b) with line driver.  
Circuit for three different types of synapses. (a) basic circuit, (b) corrected for hyperpolarisation.  
Schematic representation of an EPSP in a biological neuron (a) and in the electronic neuron (b). The dotted trace represents the timing of the pre-synaptic membrane depolarisation, but is not necessarily on the same voltage scale. The post-synaptic depolarisation is often much smaller than the pre-synaptic depolarisation.
An electronic dendrite with resistors.
Article
We present an electronic circuit modelling the spike generation process in the biological neuron. This simple circuit is capable of simulating the spiking behaviour of several different types of biological neurons. At the same time, the circuit is small so that many neurons can be implemented on a single silicon chip. This is important, as neural computation obtains its power not from a single neuron, but from the interaction between a large number of neurons. Circuits that model these interactions are also presented in this paper. They include the circuits for excitatory, inhibitory and shunting inhibitory synapses, a circuit which models the regeneration of spikes on the axon, and a circuit which models the reduction of input strength with the distance of the synapse to the cell body on the dendrite of the cell. Together these building blocks allow the implementation of electronic spiking neural networks.
 
Article
A fast prototype-based nearest neighbor classifier is introduced. The proposed Adjusted SOINN Classifier (ASC) is based on SOINN (self-organizing incremental neural network), it automatically learns the number of prototypes needed to determine the decision boundary, and learns new information without destroying old learned information. It is robust to noisy training data, and it realizes very fast classification. In the experiment, we use some artificial datasets and real-world datasets to illustrate ASC. We also compare ASC with other prototype-based classifiers with regard to its classification error, compression ratio, and speed up ratio. The results show that ASC has the best performance and it is a very efficient classifier.
 
Article
This paper proposes a generic criterion that defines the optimum number of basis functions for radial basis function (RBF) neural networks. The generalization performance of an RBF network relates to its prediction capability on independent test data. This performance gives a measure of the quality of the chosen model. An RBF network with an overly restricted basis gives poor predictions on new data, since the model has too little flexibility (yielding high bias and low variance). By contrast, an RBF network with too many basis functions also gives poor generalization performance since it is too flexible and fits too much of the noise on the training data (yielding low bias but high variance). Bias and variance are complementary quantities, and it is necessary to assign the number of basis function optimally in order to achieve the best compromise between them. In this paper we use Stein's unbiased risk estimator to derive an analytical criterion for assigning the appropriate number of basis functions. Two cases of known and unknown noise have been considered and the efficacy of this criterion in both situations is illustrated experimentally. The paper also shows an empirical comparison between this method and two well known classical methods, cross validation and the Bayesian information criterion, BIC.
 
Article
A feedforward neural network architecture aimed at survival probability estimation is presented which generalizes the standard, usually linear, models described in literature. The network builds an approximation to the survival probability of a system at a given time, conditional on the system features. The resulting model is described in a hierarchical Bayesian framework. Experiments with synthetic and real world data compare the performance of this model with the commonly used standard ones.
 
Article
Spiking neural networks (SNN) are promising artificial neural network (ANN) models as they utilise information representation as trains of spikes, that adds new dimensions of time, frequency and phase to the structure and the functionality of ANN. The current SNN models though are deterministic, that restricts their applications for large scale engineering and cognitive modelling of stochastic processes. This paper proposes a novel probabilistic spiking neuron model (pSNM) and suggests ways of building pSNN for a wide range of applications including classification, string pattern recognition and associative memory. It also extends previously published computational neurogenetic models.
 
Article
Full Text: https://edoc.hu-berlin.de/handle/18452/20143 Recent findings in neuroscience have shown differential patterns in brain activity in response to similar stimuli and activities across cultural and social differences. This calls for a framework to understand how such differences may come to be implemented in brains and neurons. Based on strands of research in social anthropology, we argue that human practices are characterized by particular patterns, and that participating in these patterns orders how people perceive and act in particular group- and context-specific ways. This then leads to a particular patterning of neuronal processes that may be detected using e.g. brain imaging methods. We illustrate this through (a) a classical example of phoneme perception (b) recent work on performance in experimental game play. We then discuss these findings in the light of predictive models of brain function. We argue that a 'culture as patterned practices' approach obviates a rigid nature-culture distinction, avoids the problems involved in conceptualizing 'culture' as a homogenous grouping variable, and suggests that participating as a competent participant in particular practices may affect both the subjective (first person) experience and (third person) objective measures of behavior and brain activity.
 
Article
A model for the study of the dynamic properties of inferior olive neuron is presented. The model, a dynamical system, comprises two autonomous components of minimal complexity that are capable of reproducing the large gamut of experimentally observed ...
 
Article
The self-organising map (SOM) has been successfully employed as a nonparametric method for dimensionality reduction and data visualisation. However, for visualisation the SOM requires a colouring scheme to imprint the distances between neurons so that the clustering and boundaries can be seen. Even though the distributions of the data and structures of the clusters are not faithfully portrayed on the map. Recently an extended SOM, called the visualisation-induced SOM (ViSOM) has been proposed to directly preserve the distance information on the map, along with the topology. The ViSOM constrains the lateral contraction forces between neurons and hence regularises the interneuron distances so that distances between neurons in the data space are in proportion to those in the map space. This paper shows that it produces a smooth and graded mesh in the data space and captures the nonlinear manifold of the data. The relationships between the ViSOM and the principal curve/surface are analysed. The ViSOM represents a discrete principal curve or surface and is a natural algorithm for obtaining principal curves/surfaces. Guidelines for applying the ViSOM constraint and setting the resolution parameter are also provided, together with experimental results and comparisons with the SOM, Sammon mapping and principal curve methods.
 
Article
Autonomous Mental Development (AMD) of robots opened a new paradigm for developing machine intelligence, using neural network type of techniques and it fundamentally changed the way an intelligent machine is developed from manual to autonomous. The work presented here is a part of SAIL (Self-Organizing Autonomous Incremental Learner) project which deals with autonomous development of humanoid robot with vision, audition, manipulation and locomotion. The major issue addressed here is the challenge of high dimensional action space (5-10) in addition to the high dimensional context space (hundreds to thousands and beyond), typically required by an AMD machine. This is the first work that studies a high dimensional (numeric) action space in conjunction with a high dimensional perception (context state) space, under the AMD mode. Two new learning algorithms, Direct Update on Direction Cosines (DUDC) and High-Dimensional Conjugate Gradient Search (HCGS), are developed, implemented and tested. The convergence properties of both the algorithms and their targeted applications are discussed. Autonomous learning of speech production under reinforcement learning is studied as an example.
 
Article
Power system stabilizers (PSS) are used to generate supplementary control signals for the excitation system in order to damp the low frequency power system oscillations. To overcome the drawbacks of conventional PSS (CPSS), numerous techniques have been proposed in the literature. Based on the analysis of existing techniques, this paper presents an indirect adaptive neural network based power system stabilizer (IDNC) design. The proposed IDNC consists of a neuro-controller, which is used to generate a supplementary control signal to the excitation system, and a neuro-identifier, which is used to model the dynamics of the power system and to adapt the neuro-controller parameters. The proposed method has the features of a simple structure, adaptivity and fast response. The proposed IDNC is evaluated on a single machine infinite bus power system under different operating conditions and disturbances to demonstrate its effectiveness and robustness.
 
Article
A key goal in the study of visual processing is to obtain a comprehensive description of the relationship between visual stimuli and neuronal responses. One way to guide the search for models is to use a general nonparametric regression algorithm, such as a neural network. We have developed a multilayer feed-forward network algorithm that can be used to characterize nonlinear stimulus-response mapping functions of neurons in primary visual cortex (area V1) using natural image stimuli. The network is capable of extracting several known V1 response properties such as: orientation and spatial frequency tuning, the spatial phase invariance of complex cells, and direction selectivity. We present details of a method for training networks and visualizing their properties. We also compare how well conventional explicit models and those developed using neural networks can predict novel responses to natural scenes.
 
Article
While artificial neural networks are regularly employed in modeling the perception of facial and vocal emotion expression as well as in automatic expression decoding by artificial agents, this approach is yet to be extended to the modeling of emotion elicitation and differentiation. In part, this may be due to the dominance of discrete and dimensional emotion models, which have not encouraged computational modeling. This situation has changed with the advent of appraisal theories of emotion and a number of attempts to develop rule-based models can be found in the literature. However, most of these models operate at a high level of conceptual abstraction and rarely include the underlying neural architecture. In this contribution, an appraisal-based emotion theory, the Component Process Model (CPM), is described that seems particularly suited to modeling with the help of artificial neural network approaches. This is due to its high degree of specificity in postulating underlying mechanisms including efferent physiological and behavioral manifestations as well as to the possibility of linking the theoretical assumptions to underlying neural architectures and dynamic processes. This paper provides a brief overview of the model, suggests constraints imposed by neural circuits, and provides examples on how the temporal unfolding of emotion can be conceptualized and experimentally tested. In addition, it is shown that the specific characteristics of emotion episodes can be profitably explored with the help of non-linear dynamic systems theory.
 
Article
Neurons in area V2 and V4 exhibit stimulus specific tuning to single stimuli, and respond at intermediate firing rates when presented with two differentially preferred stimuli ('pair response'). Selective attention to one of the two stimuli causes the neuron's firing rate to shift from the intermediate pair response towards the response to the attended stimulus as if it were presented alone. Attention to single stimuli reduces the response threshold of the neuron and increases spike synchronization at gamma frequencies. The intrinsic and network mechanisms underlying these phenomena were investigated in a multi-compartmental biophysical model of a reconstructed cat V4 neuron. Differential stimulus preference was generated through a greater ratio of excitatory to inhibitory synapses projecting from one of two input V2 populations. Feedforward inhibition and synaptic depression dynamics were critical to generating the intermediate pair response. Neuronal gain effects were simulated using gamma frequency range correlations in the feedforward excitatory and inhibitory inputs to the V4 neuron. For single preferred stimulus presentations, correlations within the inhibitory population out of phase with correlations within the excitatory input significantly reduced the response threshold of the V4 neuron. The pair response to simultaneously active preferred and non-preferred V2 populations could also undergo an increase or decrease in gain via the same mechanism, where correlations in feedforward inhibition are out of phase with gamma band correlations within the excitatory input corresponding to the attended stimulus. The results of this model predict that top-down attention may bias the V4 neuron's response using an inhibitory correlation phase shift mechanism.
 
Article
This paper introduces a generic theoretical framework for predictive learning, and relates it to data-driven and learning applications in earth and environmental sciences. The issues of data quality, selection of the error function, incorporation of the predictive learning methods into the existing modeling frameworks, expert knowledge, model uncertainty, and other application-domain specific problems are discussed. A brief overview of the papers in the Special Issue is provided, followed by discussion of open issues and directions for future research.
 
Article
The brain's most difficult computation in decision-making learning is searching for essential information related to rewards among vast multimodal inputs and then integrating it into beneficial behaviors. Contextual cues consisting of limbic, cognitive, visual, auditory, somatosensory, and motor signals need to be associated with both rewards and actions by utilizing an internal representation such as reward prediction and reward prediction error. Previous studies have suggested that a suitable brain structure for such integration is the neural circuitry associated with multiple cortico-striatal loops. However, computational exploration still remains into how the information in and around these multiple closed loops can be shared and transferred. Here, we propose a "heterarchical reinforcement learning" model, where reward prediction made by more limbic and cognitive loops is propagated to motor loops by spiral projections between the striatum and substantia nigra, assisted by cortical projections to the pedunculopontine tegmental nucleus, which sends excitatory input to the substantia nigra. The model makes several fMRI-testable predictions of brain activity during stimulus-action-reward association learning. The caudate nucleus and the cognitive cortical areas are correlated with reward prediction error, while the putamen and motor-related areas are correlated with stimulus-action-dependent reward prediction. Furthermore, a heterogeneous activity pattern within the striatum is predicted depending on learning difficulty, i.e., the anterior medial caudate nucleus will be correlated more with reward prediction error when learning becomes difficult, while the posterior putamen will be correlated more with stimulus-action-dependent reward prediction in easy learning. Our fMRI results revealed that different cortico-striatal loops are operating, as suggested by the proposed model.
 
Article
We have combined an echo state network (ESN) with a competitive state machine framework to create a classification engine called the predictive ESN classifier. We derive the expressions for training the predictive ESN classifier and show that the model was significantly more noise robust compared to a hidden Markov model in noisy speech classification experiments by 8+/-1 dB signal-to-noise ratio. The simple training algorithm and noise robustness of the predictive ESN classifier make it an attractive classification engine for automatic speech recognition.
 
Article
We analyze in this article the significance of the edge of chaos for real-time computations in neural microcircuit models consisting of spiking neurons and dynamic synapses. We find that the edge of chaos predicts quite well those values of circuit parameters that yield maximal computational performance. But obviously it makes no prediction of their computational performance for other parameter values. Therefore, we propose a new method for predicting the computational performance of neural microcircuit models. The new measure estimates directly the kernel property and the generalization capability of a neural microcircuit. We validate the proposed measure by comparing its prediction with direct evaluations of the computational performance of various neural microcircuit models. The proposed method also allows us to quantify differences in the computational performance and generalization capability of neural circuits in different dynamic regimes (UP- and DOWN-states) that have been demonstrated through intracellular recordings in vivo.
 
Article
We propose new integrated saliency map and selective motion analysis models partly inspired by a biological visual attention mechanism. The proposed models consider not only binocular stereopsis to identify a final attention area so that the system focuses ...
 
Article
We organized a challenge for IJCNN 2007 to assess the added value of prior domain knowledge in machine learning. Most commercial data mining programs accept data pre-formatted in the form of a table, with each example being encoded as a linear feature vector. Is it worth spending time incorporating domain knowledge in feature construction or algorithm design, or can off-the-shelf programs working directly on simple low-level features do better than skilled data analysts? To answer these questions, we formatted five datasets using two data representations. The participants in the "prior knowledge" track used the raw data, with full knowledge of the meaning of the data representation. Conversely, the participants in the "agnostic learning" track used a pre-formatted data table, with no knowledge of the identity of the features. The results indicate that black-box methods using relatively unsophisticated features work quite well and rapidly approach the best attainable performance. The winners on the prior knowledge track used feature extraction strategies yielding a large number of low-level features. Incorporating prior knowledge in the form of generic coding/smoothing methods to exploit regularities in data is beneficial, but incorporating actual domain knowledge in feature construction is very time consuming and seldom leads to significant improvements. The AL vs. PK challenge web site remains open for post-challenge submissions: http://www.agnostic.inf.ethz.ch/.
 
Article
Contextual modulation is a universal phenomenon in the primary visual cortex (V1). It is often allocated to the two categories of suppression and facilitation, which are either weakened or strengthened by the contextual stimuli, respectively. A number of experiments in neurophysiology have elucidated their important functions in visual information processing, such as contour integration, figure-ground segregation, saliency map and so on. A computational model, inspired by visual cortical mechanisms of contextual modulation, is presented in this paper. We first give separate models for surround suppression (SS), collinear facilitation (CF) and cross-orientation facilitation (COF), respectively, then unify them to a mixed model. Model behavior has then been tested using synthetical images and nature images, and is consistent with the data of physiological experimentation. We achieve fine results using the model to extract salient structures and contours from images. This work develops a computational model, using the perceptual mechanisms in V1, and provides a biologically plausible strategy for computer vision.
 
Article
The problem of controlling locomotion is an area in which neuroscience and robotics can fruitfully interact. In this article, I will review research carried out on locomotor central pattern generators (CPGs), i.e. neural circuits capable of producing coordinated patterns of high-dimensional rhythmic output signals while receiving only simple, low-dimensional, input signals. The review will first cover neurobiological observations concerning locomotor CPGs and their numerical modelling, with a special focus on vertebrates. It will then cover how CPG models implemented as neural networks or systems of coupled oscillators can be used in robotics for controlling the locomotion of articulated robots. The review also presents how robots can be used as scientific tools to obtain a better understanding of the functioning of biological CPGs. Finally, various methods for designing CPGs to control specific modes of locomotion will be briefly reviewed. In this process, I will discuss different types of CPG models, the pros and cons of using CPGs with robots, and the pros and cons of using robots as scientific tools. Open research topics both in biology and in robotics will also be discussed.
 
Article
The striatum, the principal input structure of the basal ganglia, is crucial to both motor control and learning. It receives convergent input from all over the neocortex, hippocampal formation, amygdala and thalamus, and is the primary recipient of dopamine in the brain. Within the striatum is a GABAergic microcircuit that acts upon these inputs, formed by the dominant medium-spiny projection neurons (MSNs) and fast-spiking interneurons (FSIs). There has been little progress in understanding the computations it performs, hampered by the non-laminar structure that prevents identification of a repeating canonical microcircuit. We here begin the identification of potential dynamically-defined computational elements within the striatum. We construct a new three-dimensional model of the striatal microcircuit's connectivity, and instantiate this with our dopamine-modulated neuron models of the MSNs and FSIs. A new model of gap junctions between the FSIs is introduced and tuned to experimental data. We introduce a novel multiple spike-train analysis method, and apply this to the outputs of the model to find groups of synchronised neurons at multiple time-scales. We find that, with realistic in vivo background input, small assemblies of synchronised MSNs spontaneously appear, consistent with experimental observations, and that the number of assemblies and the time-scale of synchronisation is strongly dependent on the simulated concentration of dopamine. We also show that feed-forward inhibition from the FSIs counter-intuitively increases the firing rate of the MSNs. Such small cell assemblies forming spontaneously only in the absence of dopamine may contribute to motor control problems seen in humans and animals following a loss of dopamine cells.
 
Article
Evolution of neural networks, or neuroevolution, has been a successful approach to many low-level control problems such as pole balancing, vehicle control, and collision warning. However, certain types of problems-such as those involving strategic decision-making-have remained difficult for neuroevolution to solve. This paper evaluates the hypothesis that such problems are difficult because they are fractured: The correct action varies discontinuously as the agent moves from state to state. A method for measuring fracture using the concept of function variation is proposed and, based on this concept, two methods for dealing with fracture are examined: neurons with local receptive fields, and refinement based on a cascaded network architecture. Experiments in several benchmark domains are performed to evaluate how different levels of fracture affect the performance of neuroevolution methods, demonstrating that these two modifications improve performance significantly. These results form a promising starting point for expanding neuroevolution to strategic tasks.
 
Article
How does an infant learn through visual experience to imitate actions of adult teachers, despite the fact that the infant and adult view one another and the world from different perspectives? To accomplish this, an infant needs to learn how to share joint attention with adult teachers and to follow their gaze towards valued goal objects. The infant also needs to be capable of view-invariant object learning and recognition whereby it can carry out goal-directed behaviors, such as the use of tools, using different object views than the ones that its teachers use. Such capabilities are often attributed to "mirror neurons". This attribution does not, however, explain the brain processes whereby these competences arise. This article describes the CRIB (Circular Reactions for Imitative Behavior) neural model of how the brain achieves these goals through inter-personal circular reactions. Inter-personal circular reactions generalize the intra-personal circular reactions of Piaget, which clarify how infants learn from their own babbled arm movements and reactive eye movements how to carry out volitional reaches, with or without tools, towards valued goal objects. The article proposes how intra-personal circular reactions create a foundation for inter-personal circular reactions when infants and other learners interact with external teachers in space. Both types of circular reactions involve learned coordinate transformations between body-centered arm movement commands and retinotopic visual feedback, and coordination of processes within and between the What and Where cortical processing streams. Specific breakdowns of model processes generate formal symptoms similar to clinical symptoms of autism.
 
Article
Gene-expression microarray is a novel technology that allows the examination of tens of thousands of genes at a time. For this reason, manual observation is not feasible and machine learning methods are progressing to face these new data. Specifically, ...
 
Five different states of the grid model representing five different events of a simple dynamic simulation in which first source (top left) is active from [0, 350] ms,  
Comparison of two similar normal rhythm cases. Time series plots (top) look very similar while trajectory plots (bottom) show noticeable differences (19 600 points).  
Article
Neurodynamical systems are characterized by a large number of signal streams, measuring activity of individual neurons, local field potentials, aggregated electrical (EEG) or magnetic potentials (MEG), oxygen use (fMRI) or activity of simulated neurons. Various basis set decomposition techniques are used to analyze such signals, trying to discover components that carry meaningful information, but these techniques tell us little about the global activity of the whole system. A novel technique called Fuzzy Symbolic Dynamics (FSD) is introduced to help in understanding of the multidimensional dynamical system's behavior. It is based on a fuzzy partitioning of the signal space that defines a non-linear mapping of the system's trajectory to the low-dimensional space of membership function activations. This allows for visualization of the trajectory showing various aspects of observed signals that may be difficult to discover looking at individual components, or to notice otherwise. FSD mapping can be applied to raw signals, transformed signals (for example, ICA components), or to signals defined in the time-frequency domain. To illustrate the method two FSD visualizations are presented: a model system with artificial radial oscillatory sources, and the output layer (50 neurons) of Respiratory Rhythm Generator (RRG) composed of 300 spiking neurons.
 
Article
Targeted patch-clamp recordings are a promising technique that can directly address the physiological properties of a specific neuron embedded in a neuronal network. Typically, neurons are visualized through fluorescent dyes or fluorescent proteins with ...
 
Article
Brains are complex networks. Previously, we revealed that specific connected structures are either significantly abundant or rare in cortical networks. However, it remains unknown whether systems from other disciplines have similar architectures to brains. ...
 
Article
We organized a challenge in "Unsupervised and Transfer Learning": the UTL challenge (http://clopinet.com/ul). We made available large datasets from various application domains: handwriting recognition, image recognition, video processing, text processing, and ecology. The goal was to learn data representations that capture regularities of an input space for re-use across tasks. The representations were evaluated on supervised learning "target tasks" unknown to the participants. The first phase of the challenge was dedicated to "unsupervised transfer learning" (the competitors were given only unlabeled data). The second phase was dedicated to "cross-task transfer learning" (the competitors were provided with a limited amount of labeled data from "source tasks", distinct from the "target tasks"). The analysis indicates that learned data representations yield significantly better results than those obtained with original data or data preprocessed with standard normalizations and functional transforms.
 
Article
Artificial neural networks increasingly involve spiking dynamics to permit greater computational efficiency. This becomes especially attractive for on-chip implementation using dedicated neuromorphic hardware. However, both spiking neural networks and ...
 
Article
We modeled and analyzed a signal transduction system of long-term potentiation (LTP) in hippocampal post-synapse. Bhalla and Iyengar [Science 283(1999) 381] have developed a hippocampal LTP model. In the conventional model, the concentration of protein phosphatase 2A (PP2A) was fixed. However, it was reported that dynamic inactivation of PP2A was essential for LTP [J. Neurochem. 74 (2000) 807]. We introduced a dynamic modeling of PP2A; inactivation (phosphorylation) of PP2A by calcium/calmodulin-dependent protein kinase II (CaMKII) in the presence of calcium/calmodulin, self-activation (autodephosphorylation) of PP2A, and inactivation (dephosphorylation) of CaMKII by PP2A. This model includes complex feedback loops; both CaMKII and PP2A are autoactivated, while they inactivate each other. Moreover, we proposed an analysis strategy for model validation by applying the results of sensitivity analysis. In our system, calcineurin (CaN) played an essential role, rather than the activation of protein kinase C (PKC) as documented in the conventional model. From results of the analysis of our model, we found the following robustness as characteristics of bistability in our model: (1). PP2A reactions against calcium ion (Ca(2+)) perturbation; (2). PP2A inactivation against PP2A increase; (3). protein phosphatase 1 (PP1) activation against PF2A increase; and (4). PP2A reactions against PP2A initial concentration. These properties facilitated LTP induction in our system. We showed that another mechanism could introduce bistable behavior by adding dynamic reactions of PP2A.
 
Article
Although sigma-delta modulation is widely used for analog-to-digital (A/D) converters, sigma-delta concepts are only for 1D signals. Signal processing in the digital domain is extremely useful for 2D signals such as used in image processing, medical imaging, ultrasound imaging, and so on. The intricate task that provides true 2D sigma-delta modulation is feasible in the spatial domain sigma-delta modulation using the discrete-time cellular neural network (DT-CNN) with a C-template. In the proposed architecture, the A-template is used for a digital-to-analog converter (DAC), the C-template works as an integrator, and the nonlinear output function is used for the bilevel output. In addition, due to the cellular neural network (CNN) characteristics, each pixel of an image corresponds to a cell of a CNN, and each cell is connected spatially by the A-template. Therefore, the proposed system can be thought of as a very large-scale and super-parallel sigma-delta modulator. Moreover, the spatio-temporal dynamics is designed to obtain an optimal reconstruction signal. The experimental results show the excellent reconstruction performance and capabilities of the CNN as a sigma-delta modulator.
 
Article
In order to plan accurate motor actions, the brain needs to build an integrated spatial representation associated with visual stimuli and haptic stimuli. Since visual stimuli are represented in retina-centered co-ordinates and haptic stimuli are represented in body-centered co-ordinates, co-ordinate transformations must occur between the retina-centered co-ordinates and body-centered co-ordinates. A spiking neural network (SNN) model, which is trained with spike-timing-dependent-plasticity (STDP), is proposed to perform a 2D co-ordinate transformation of the polar representation of an arm position to a Cartesian representation, to create a virtual image map of a haptic input. Through the visual pathway, a position signal corresponding to the haptic input is used to train the SNN with STDP synapses such that after learning the SNN can perform the co-ordinate transformation to generate a representation of the haptic input with the same co-ordinates as a visual image. The model can be applied to explain co-ordinate transformation in spiking neuron based systems. The principle can be used in artificial intelligent systems to process complex co-ordinate transformations represented by biological stimuli.
 
Article
In the tasks of image representation, recognition and retrieval, a 2D image is usually transformed into a 1D long vector and modelled as a point in a high-dimensional vector space. This vector-space model brings up much convenience and many advantages. However, it also leads to some problems such as the Curse of Dimensionality dilemma and Small Sample Size problem, and thus produces us a series of challenges, for example, how to deal with the problem of numerical instability in image recognition, how to improve the accuracy and meantime to lower down the computational complexity and storage requirement in image retrieval, and how to enhance the image quality and meanwhile to reduce the transmission time in image transmission, etc. In this paper, these problems are solved, to some extent, by the proposed Generalized 2D Principal Component Analysis (G2DPCA). G2DPCA overcomes the limitations of the recently proposed 2DPCA (Yang et al., 2004) from the following aspects: (1) the essence of 2DPCA is clarified and the theoretical proof why 2DPCA is better than Principal Component Analysis (PCA) is given; (2) 2DPCA often needs much more coefficients than PCA in representing an image. In this work, a Bilateral-projection-based 2DPCA (B2DPCA) is proposed to remedy this drawback; (3) a Kernel-based 2DPCA (K2DPCA) scheme is developed and the relationship between K2DPCA and KPCA (Scholkopf et al., 1998) is explored. Experimental results in face image representation and recognition show the excellent performance of G2DPCA.
 
Article
Robust dimensionality reduction is an important issue in processing multivariate data. Two-dimensional principal component analysis based on L1-norm (2DPCA-L1) is a recently developed technique for robust dimensionality reduction in the image domain. The basis vectors of 2DPCA-L1, however, are still dense. It is beneficial to perform a sparse modelling for the image analysis. In this paper, we propose a new dimensionality reduction method, referred to as 2DPCA-L1 with sparsity (2DPCAL1-S), which effectively combines the robustness of 2DPCA-L1 and the sparsity-inducing lasso regularization. It is a sparse variant of 2DPCA-L1 for unsupervised learning. We elaborately design an iterative algorithm to compute the basis vectors of 2DPCAL1-S. The experiments on image data sets confirm the effectiveness of the proposed approach.
 
Top-cited authors
Maxwell Stinchcombe
  • University of Texas at Austin
Erkki Oja
  • Aalto University
David H. Wolpert
  • Santa Fe Institute
Mitsuo Kawato
  • Advanced Telecommunications Research Institute
Alex Graves
  • University of Toronto