Advances in Neuro-Information Processing

Integrative evolving connectionist systems (iECOS) integrate principles from different levels of information processing in the brain, including cognitive-, neuronal-, genetic- and quantum, in their dynamic interaction over time. The paper introduces a new framework of iECOS called integrative probabilistic evolving spiking neural networks (ipSNN) that incorporate probability learning parameters. ipSNN utilize a quantum inspired evolutionary optimization algorithm to optimize the probability parameters as these algorithms belong to the class of estimation of distribution algorithms (EDA). Both spikes and input features in ipESNN are represented as quantum bits being in a superposition of two states (1 and 0) defined by a probability density function. This representation allows for the state of an entire ipESNN at any time to be represented probabilistically in a quantum bit register and probabilistically optimised until convergence using quantum gate operators and a fitness function. The proposed ipESNN is a promising framework for both engineering applications and brain data modeling as it offers faster and more efficient feature selection and model optimization in a large dimensional space in addition to revealing new knowledge that is not possible to obtain using other models. Further development of ipESNN are the neuro-genetic models – ipESNG, that are introduced too, along with open research questions.

In this paper, we construct a spiking network model based on the firing-rate coding hippocampal model proposed by Becker. Basal training patterns are presented to the model network and spiking self organizing map learning is applied to the network in order to store the training patterns. We then apply a morphogenesis model in the dentate gyrus region to generate new neurons and investigate the influence of such neurogenesis on the storage and recall of novel memory. As a result, the storage capacity is essentially unchanged by the morphogenetic algorithm even when the number of training patterns is changed.

The framework of neuroevolution (NE) provides a way of finding appropriate structures as well as connection weights of artificial neural networks. However, the conventional NE approach of directly coding each connection weight by a gene is severely limited in its scalability and evolvability. In this study, we propose a novel indirect coding approach in which a phenotypical network develops from the genes encoding multiple subnetwork modules. Each gene encodes a subnetwork consisting of the input, hidden, and output nodes and connections between them. A connection can be a real weight or a pointer to another subnetwork. The structure of the network evolves by inserting new connection weights or subnetworks, merging two subnetworks as a higher-level subnetwork, or changing the existing connections. We investigated the evolutionary process of the network structure using the task of double pole balancing. We confirmed that the proposed method by the modular developmental rule can produce a wide variety of network architectures and that evolution can trim them down to the most appropriate ones required by the task.

Neural Pattern-Based Reasoning for real-world robot navigation problems is proposed. Based on the concept of Pattern-Based Reasoning, the method enables a mobile robot to solve goal-oriented navigation problems in a remarkably short time with low memory consumption. Given a simple learning environment, the observed input vectors are processed by a Self-Organizing Incremental Neural Network (SOINN) to generate Spatial Common Patterns (CPs), which are useful in other unfamiliar environments. Performing goal-oriented navigation in unfamiliar environments, with prior information neither of the map nor the goal, the robot recognizes the partial area by referring to the nearest CPs and forming a pattern of CPs called A-Pattern. The sequential A-Patterns are used to derive the map of the environment. This map is optimized based on reasoning, as the new transitions between areas could be generated automatically. The method is evaluated by solving one real-world maze and three Webots simulated mazes. The results show that the proposed method enables the robot to find the markedly shorter path in only one episode, whereas use of the Reinforcement Learning requires more episodes. The map contains more information than the current hybrid map building or topological map. The map does not rely on coordinate, resulting in non-sensitivity to the error in self-pose estimation.

The amygdala is the neural structure that acts as an evaluator of potentially threatening stimuli. We present a biologically plausible model of the visual fear conditioning pathways leading to the amygdala, using a topographic conditioning map (TCM). To evaluate the model, we first use abstract stimuli to understand its ability to form topographic representations, and subsequently to condition on arbitrary stimuli. We then present results on facial emotion recognition using the sub-cortical pathway of the model. Compared to other emotion classification approaches, our model performs well, but does not have the need to pre-specify features. This generic ability to organise visual stimuli is enhanced through conditioning, which also improves classification performance. Our approach demonstrates that a biologically motivated model can be applied to real-world tasks, while allowing us to explore biological hypotheses.

This paper presents a model for the development of primitive nervous systems in a hydra-like animat controlled by a gene regulatory network. The gene regulatory network consists of structural genes that simulate the main cellular events during neural development at an abstract level, namely, cell division, cell migration, and axon growth, and regulatory genes that control the expression of the structural genes. The developmental model is evolved with an evolutionary algorithm to achieve the correct developmental order. After the genetically controlled neural development is completed, the connectivity and weights of the neural networks are further adapted to allow the animat for performing simple behaviors such as the food catching behavior of a hydra. Our preliminary results suggest that the proposed developmental model is promising for computational simulation of the evolution of neural development for understanding neural organization in biological organisms.

In this paper it is shown that Reservoir Computing can be successfully applied to perform real-time detection of epileptic seizures in Electroencephalograms (EEGs). Absence and tonic-clonic seizures are detected on intracranial EEG coming from rats. This resulted in an area under the Receiver Operating Characteristics (ROC) curve of about 0.99 on the data that was used. For absences an average detection delay of 0.3s was noted, for tonic-clonic seizures this was 1.5s. Since it was possible to process 15h of data on an average computer in 14.5 minutes all conditions are met for a fast and reliable real-time detection system.

Subgoal learning is investigated to effectively build a goal-oriented behavior control rule with which a mobile robot can achieve a task goal for any starting task configurations. For this, states of interest are firstly extracted from successful task episodes, where the averaged occurrence frequency of states is used as threshold value to identify states of interest. And, subgoals are learned by clustering similar features of state transition tuples. Here, features used in clustering are produced by using changes of the states in the state transition tuples. A goal-oriented behavior control rule is made in such a way that proper actions are sequentially and/or reactively generated from the subgoal according to the context of states. To show the validities of our proposed subgoal learning as well as a goal-oriented control rule of mobile robots, a Box-Pushing-Into-a-Goal(

BPIG

) task is simulated and experimented.

The following coding mechanisms in the CA3-CA1 hippocampal networks were examined. Firstly, the way in which the information of the spatio-temporal sequence of the hippocampal CA1 pyramidal cells was clarified by using the patch-clamp recording method. The input-output relations were analyzed by applying “spatial clustering index” and its “self-similarity” (Cantor-like coding) measure of the sequences. The membrane potentials were hierarchically clustered in a self-similar manner to the input sequences. The property was found to be present at one and two time steps retrograde in the sequences. The experimental results closely matched theoretical results of Cantor coding, reported by Tsuda and Kuroda (2001). Secondly, in the consolidation process, the spatiotemporal learning rule (STLR) composed of the spatial coincidence and its time history plays an important role in mapping the Cantor-like property onto synaptic weight space. The coexistence of STLR and Cantor-like coding in single pyramidal neuron of the hippocampal CA1 area is discussed from the viewpoint of coding mechanisms of reinforcement learning.

Engineering approaches to stereo typically use explicit search for the best matching between left and right sub-windows, which involves a high cost of search and unstable performance in the presence of binocular inconsistency and weak texture. The brain does not seem to conduct explicit search in the V1 and V2 cortex. But the mechanisms that the brain employs to integrate binocular disparity into 3-D perception is still largely a mystery. The work presented in this paper focuses on an important issue of integrated stereo: How the same cortex can perform recognition and perception by generating a topographic disparity-tuning map using top-down connections. As top-down connections with object-class supervisory signals result in topographic class maps, the model presented here clarifies that stereo can be processed by a unified in-place learning framework in the neural layers, and can generate iconic-abstract internal representation.

We predicted the rate constants of AMPA receptor channels kinetics models of neural cells which could induce the change in efficacy of synaptic transmission by computer simulation. Excitatory postsynaptic currents (EPSCs) were reconstructed by computer calculation with the proposed kinetics models. Moreover, electrical responses were measured from the cells by using 2-photon laser uncaging. It was shown that the properties of the evoked current responses by photolysis and those obtained from spontaneous synaptic currents have the similar properties which indicates that the present AMPA receptor channel models mediate the evoked responses by laser photolysis We investigated and proposed the possible rate constants, which could explain the changes in the EPSCs amplitude during LTP/LTD without changing the waveform, and corresponding physiological elements to them. Moreover, it is suggested that the present method based on kinetics models could be used for the investigation of other experiments evoked by laser photolysis.

This paper shows that a biology-based neural network (called a central pattern generator (CPG)) can be re-evolved for sea conditions / frequencies. The fish’s CPG operates at 1.74Hz to 5.56Hz, whereas we require performance to reach 0.05Hz to 0.35Hz (20s to 3s waves) for an alternative engineering problem. This is to enable adaptive control of wave energy devices, increasing their efficiency and power yield. To our knowledge, this is the first time a bio-inspired circuit will be integrated into the engineering domain (and for a completely different function). This provides great inspiration for utilising other neural network mechanisms for alternative tasks.

One of the fundamental motivations for feature selection is to overcome the curse of dimensionality. A novel feature selection algorithm is developed in this chapter based on a combination of Differential Evolution (DE) optimization technique and statistical feature distribution measures. The new algorithm, referred to as DEFS, utilizes the DE float number optimizer in a combinatorial optimization problem like feature selection. The proposed DEFS highly reduces the computational cost while at the same time proves to present a powerful performance. The DEFS is tested as a search procedure on different datasets with varying dimensionality. Practical results indicate the significance of the proposed DEFS in terms of solutions optimality and memory requirements.

We have extracted brain functional networks from fMRI data based on temporal correlations of voxel activities during the rest and task periods. The goal of our preliminary research was to study the topology of these networks in terms of small-world and scale-free properties. The small-world property was quite clearly evident whereas the scale-free character was less obvious, especially in the rest condition. In addition, there were some differences between the rest and task functional brain networks as well as between subjects. We discuss the relation of properties of functional brain networks to the topological properties of the underlying anatomical networks, which are largely dependent upon genetic instructions during brain development.

One of most important concepts in robotics and artificial intelligence is the embodied approach, focusing on the importance of having a body that functionally connects to the external world. This setup suggests that the intelligence develops through sensorimotor skills and through situations that would actually be confronted in the environment. We support this concept and propose to further extend it to embodiment in the time domain. Nervous systems have variable processing times. The different time courses proceed in the nervous system in parallel, and individual circuits independently and cooperatively work under the constraints of temporal properties. We here propose an experimental platform of oscillatory neural networks having real-time communication with the environment through the robot’s body. The synchronization mechanism of oscillations in neural activities have the advantage of synthetic controls known in motor coordination, but we extend this to circuits for cognitive functions like episodic memory formation and decision making of the robotic behavior by using the theta phase coding mechanism. A slow oscillation, like the theta rhythm, enables behavioral temporal sequences to be compressed in sequential firings during each oscillation cycle, and this helps to represent cognitive information in episodes composed of past-present-future structures. The temporal structure is crucial for recognition of the current context and adaptability in dynamic environments, and it smoothly controls sensorimotor local circuits with faster time scales. This work represents a tiny step towards constructing the brain by focusing on the temporal structure, yet this approach may elucidate the new nature of the brain-based intelligence.

First-spiking dynamics of optimally controlled neuron under stimulation of colored noise is investigated. The stochastic averaging principle is utilized and the model equation is approximated by diffusion process and depicted by Itô stochastic differential equation. The control problems for maximizing the resting probability and maximizing the time to first spike are constructed and the dynamical programming equations associated with the corresponding optimization problem are established. The optimal control law is determined. The corresponding backward Kolmogorov equation and Pontryagin equation are established and solved to yield the resting probability and the time to first spike. The analytical results are verified by Monte Carlo simulation. It has shown that the proposed control strategy can suppress the overactive neuronal firing activity and possesses potential application for some neural diseases treatment.

The Bienenstock-Cooper-Munroe (BCM) rule is one of the best-established learning formalisms for neural tissue. However, as it is based on pulse rates, it can not account for recent spike-based experimental protocols that have led to spike timing dependent plasticity (STDP) rules. At the same time, STDP is being challenged by experiments exhibiting more complex timing rules (e.g. triplets) as well as simultaneous rate- and timing dependent plasticity. We derive a formulation of the BCM rule which is based on the instantaneous postsynaptic membrane potential as well as the transmission profile of the presynaptic spike. While this rule is neither directly rate nor timing based, it can replicate BCM, conventional STDP and spike triplet experimental data, despite incorporating only two state variables. Moreover, these behaviors can be replicated with the same set of only four free parameters, avoiding the overfitting problem of more involved plasticity rules.

In this paper, we propose a hybrid spiking neuron which can exhibit various bifurcation phenomena and response characteristics of inter spike intervals. Using a discrete/continuous-states hybrid map, we can clarify typical bifurcation mechanisms and can analyze the response characteristics. In addition, we propose a learning algorithm of the hybrid spiking neuron and show that the neuron can approximate given response characteristics of inter spike intervals.

Event-related potentials (ERP) reflect external/internal stimuli. It is of our interest in the current study to investigate ERPs from spontaneous motor actions. Spontaneous motor actions are voluntary motor tasks carried out at an individual’s own pace. Various properties of spontaneous motor actions are revealed through event-related desynchronisation and event-related synchronisation. These properties are reported and discussed in this paper.

We investigated competition between synapses located proximal and distal dendrites through STDP learning rules using a dentate granule cell model. The proximal dendrite and the distal dendrite were stimulated by a regular pulse train and a random pulse train respectively. When both synapses were subject to asymmetric STDP rules, the distal synapse was not enhanced but the proximal synapse was enhanced through synaptic competition. This competition was caused when the integral of the LTD window was larger than that of the LTP window of the STDP rule in the distal dendrite. In contrast, when the proximal synapse was not subject to the asymmetric STDP rule but a Mexican-hat STDP rule, location of the synapse enhanced by synaptic competition was switched from the proximal synapse to the distal synapse. We also examined the role of inhibitory interneurons in the synaptic competition.

Herein, we make a theoretical effort to characterize the interplay of the main stimuli underlying the cardiac control. Based on the analysis of heartbeat intervals and using neural coding strategies, we investigate the hypothesis that information theoretic principles could be used to give insights to the strategy evolved to control the heart. This encodes the sympathetic and parasympathetic stimuli. As a result of analysis, we illustrate and emphasize the basic sources that might be attributed to control the heart rate based on the interplay of the autonomic tones.

A variety of (dis)similarity measures for one-dimensional point processes (e.g., spike trains) are investigated, including the Victor-Purpura distance metric, the van Rossum distance metric, the Schreiber

et al.

similarity measure, the Hunter-Milton similarity measure, the event synchronization proposed by Quiroga, and the stochastic event synchrony measures (SES) recently proposed by Dauwels

et al.

By analyzing surrogate data, it is demonstrated that most measures are not able to distinguish timing precision and event reliability, i.e., they depend on both aspects of synchrony. There are two exceptions: with appropriate choice of parameters, event synchronization quantifies event reliability, independently of timing precision; the two SES parameters quantify both timing precision and event reliability separately. Before one can apply the (dis)similarity measures (with the exception of SES), one needs to determine potential lags between the point processes. On the other hand, SES deals with lags in a natural and direct way, and therefore, the SES similarity measures are robust to lags.

As an illustration, neuronal spike data generated by the Morris-Lecar neuron model is considered.

Neural dynamics of the ”theta phase precession” in the hippocampus are known to have a computational advantage with respect to memory encoding. Computational studies have shown that a combination of theta phase precession and spike-timing-dependent plasticity (STDP) can serve as recurrent networks in various methods of memory storage. Conversely, the proposed dynamics of neurons and synapses appear too complicated to give any clear perspective on the network formation in the case of a large number of neurons (>1000). In this paper, we theoretically analyzed the evolution of synaptic weights under a given input sequence. We present our results as a simple equation demonstrating that the magnitude of the slow component of an input sequence giving successive coactivation results in asymmetric connection weights. Further comparison with computer experiments confirms the predictability of network formation.

This paper proposes that spectral characteristics of background neuronal potentials can be effective parameters to classifying and identifying neural activities from subthalamic nucleus (STN) and subtantia nigra (SNr). The spike-free background signals were obtained from inter-spike microelectrode recording signals. The averaged periodogram was then used to compute the power spectral density of the background signals. Three spectral parameters were extracted and used as discriminant features for artificial neural networks. The commonly used neuronal firing patterns were also extracted from the detected neuronal spikes and used as discriminant features. Our results showed that the classification performance based on background parameters was similar or better than using neuronal firing patterns. This implied that neuronal background can be useful as an aid in targeting STN as well as neuronal firing patterns, saving from spike identification as single- or multi-neuron discharges.

“Echo State” neural networks, which are a special case of recurrent neural networks, are studied from the viewpoint of their learning ability, with a goal to achieve their greater predictive ability. In this paper we study the influence of the memory length on predictive abilities of Echo State neural networks. The conclusion is that Echo State neural networks with fixed memory length can have troubles with adaptation of its intrinsic dynamics to dynamics of the prediction task. Therefore, we have tried to create complex prediction system as a combination of the local expert Echo State neural networks with different memory length and one special gating Echo State neural network. This approach was tested in laser fluctuations prediction. The prediction error achieved by this approach was substantially smaller in comparison with prediction error achieved by standard Echo State neural networks.

In the mammalian inner ear, the inner hair cell transforms a sound-induced mechanical vibration into an electric potential. The spiral ganglion cell encodes the electric potential into a spike-train which is transmitted to the central nervous system. In this paper we present a novel artificial electrical circuit model of the spiral ganglion cell. We derive a return map which can analytically describe dynamics of the model. Using the map, we can derive theorems that guarantee that the presented model can realize some of important properties of the biological spiral ganglion cell. The theorems are confirmed numerically.

The role of efficient information processing in organizing nervous systems is investigated. For this purpose, we have developed a computational model termed the

Hydramat Simulation Environment

, so named since it simulates certain structural aspects of fresh water hydra. We compare the evolution of neural organization in architectures that remain static throughout their lifetimes and neural architectures that are perturbed by small random amounts. We find that (a) efficient information processing directly contributes to the structural organization of a model nervous system and (b) lifetime architectural perturbations can facilitate novel architectural features.

Bump modeling is a method used to extract oscillatory bursts in electrophysiological signals, who are most likely to be representative of local synchronies. In this paper we present an improved sparse bump modeling method. The improvements are done in the adaptation method by optimizing the parameters according to the order of their derivatives; and in the window matching method by changing the selection of the initial function. Experimental results, comparing previous method

vs

the improved version, show that the obtained model fits better the signal, hence the result will be much more precise and useful.

Motor related potentials are generated when an individual is engaged in a task involving motor actions. The transient post-synaptical potential could be observed from the recorded electroencephalogram (EEG) signal. Properties derived from time domain and frequency domain such as event-related motor potential and suppression in band power could be useful EEG features. In this report,

lateralised motor potential (LMP)

and

band power ratio (BPR)

are used to classify cued left-fingers and right-fingers movements. Two classifiers are employed in this experiment: minimum distance classifier (MDC) and normal density Bayes classifier (NDBC). The results show that the features from LMP has more discriminative power than band power ratio. They also show that NDBC has a perfect performance in this task.

This paper presents a neural network based hierarchical motor schema of a multi finger hand to generate suitable behavior for an unknown situation without retraining all neural networks and investigates its motion diversity by changing its input signals. Conventional neural networks are hard to generate desired movements in an unknown situation. Our hierarchical motor schema consists of the two layers. A lower schema is implemented by a recurrent neural network trained with primitive movement patterns and generates a finger movement from a command code sent from the upper schema. The upper schema generates command codes to each finger from a behavior command code such as grasping. We showed that though the lower schemata were fixed, diversity of generated finger movements can be obtained by changing a behavior code of the upper schema through computer simulation.

We provide a biological justification for the success of spectral domain models of visual attention and propose a refined spectral domain based spatiotemporal saliency map model including a more biologically plausible method for motion saliency generation. We base our approach on the idea of spectral whitening (SW), and show that this whitening process is an estimation of divisive normalization, a model of lateral surround inhibition. Experimental results reveal that SW is a better performer at predicating eye fixation locations than other state-of-the-art spatial domain models for color images, achieving a 92% consistency with human behavior in urban environments. In addition, the model is simple and fast, capable of generating saliency maps in real-time.

Evaluation of global closure of contour is important in object perception in natural scenes including many occlusions. Here, we conducted psychophysical experiments to test the closure effect in contour integration within a single spatial frequency range using Gabor patches. We found that closed arrangements of the patches embedded in noise patches were more salient than open arrangements. This effect was seen even in the experiment using stimuli that isolated global closure effect excluding effect of local orientational continuity among adjacent patches. These findings suggest that processing for global closure of contour exists separately from that for local orientational continuity in the visual system.

We present a new analysis on the heuristic strategy developed in the hippocampal circuit through the memory and learning process. A heuristic approach rapidly leads a solution close to the best possible answer utilizing easy-access information under the situation in which it is difficult to find the best answer. Focusing on the day trading, which needs the rapid decision making within a restricted time, we demonstrate that the heuristic strategy emerges in the process of the memory integration through the compensation for the limit of the information processing ability of the brain. We expect that findings from our trials will help to reveal the hippocampal role on the establishment of decision making strategies and provide the new idea in order to predict the social behavior or improve the current computer power.

Tracking an object in depth is an important task, since the distance to an object often correlates with an imminent danger, e.g. in the case of an approaching vehicle. A common way to estimate the depth of a tracked object is to utilize binocular methods like stereo disparity. In practice, however, depth measurement using binocular methods is technically expensive due to the need of camera calibration and rectification. In addition, higher depths are difficult to estimate because of an inverse relationship between disparity and depth. In this paper a new approach for depth estimation, depth-from-sizes (DFS), is introduced. We present a human-inspired monocular method where the depth, the physical size and the retinal size of the object are estimated in a mutually interdependent manner. For each of the three terms specific measurement and estimation methods are probabilistically combined. In two evaluation scenarios it is shown that this approach is a reliable alternative to the standard stereo disparity approach for depth estimation with several advantages: 1) simultaneous estimation of depth, physical size and retinal size; 2) no stereo camera calibration and rectification; 3) good depth estimation at higher depth ranges.

This work provide a short study of training algorithms useful for adaptation of recurrent connectionist models for symbolic time series modeling tasks. We show that approaches based on Kalman filtration outperform standard gradinet based training algorithms. We propose simple approximation to the Kalman filtration with favorable computational requirements and on several linguistic time series taken from recently published papers we demonstrate superior ability of the proposed method.

A computational model of a car driver’s cognitive process was developed to create more useful assistance technology for making driving safer. The model’s status is mainly determined by the driver’s eye movement. In this model, we defined two types of risk : explicit risk for visible objects and inexplicit risk for non-visible areas. In our simulation, we attempted to reconstruct the driver’s eye movement while driving using our model.

There is little understanding on representation and reconstruction of object shape in the cortex. Physiological studies with macaque suggested that neurons in V1 respond to Medial-Axis (MA). We investigated whether (1) early visual areas could provide basis for MA representation, and (2) we could reconstruct the original shape from its MA, with a physiologically realistic computational model consisting of early- to intermediate-level visual areas. Assuming the synchronization of border-ownership selective cells at stimulus onset, our model was capable of detecting MA, indicating that early visual area could provide basis for MA representation. Furthermore, we propose a physiologically plausible reconstruction algorithm with the summation of distinct gaussians.

The correlation-based rule of plasticity has been widely believed to be involved in the organization of functional synaptic circuits. However, recent studies have suggested that the direction of plasticity in the sensory-deprived barrel cortex can be reversed, depending on the stimulus environment, from that predicted by the correlation-based plasticity. To investigate whether spike-timing-dependent plasticity (STDP) may underlie such reversal in cortical plasticity, we study the influence of the correlation time on the synaptic cooperative and competitive mechanisms based on the input correlation. The results show that in the presence of activity-dependent feedback modification of the STDP window function, the increase in the correlation time can reverse the plasticity outcome such that for shorter correlation time, more frequently activated synapses are strengthened while, as the correlation time is sufficiently prolonged, less frequently activated synapses become strengthened.

A method is proposed to determine the similarity of a collection of time series. As a first step, one extracts events from the time series, in other words, one converts each time series into a point process (“event sequence”); next one tries to align the events from those different point processes. The better the events can be aligned, the more similar the original time series are considered to be. The proposed method is applied to predict mild cognitive impairment (MCI) from EEG and to investigate the dynamics of oscillatory-event synchrony of steady-state visually evoked potentials (SSVEP).

Primates can use tools effectively, and each of the species exhibits different tool-use behavior. We study what difference in tool-use behavior will be observed if different internal models of the environment are used for planning. We compare two agents with different internal models, each of which consists of an artificial neural network. The first agent (agent 1) has an internal model that predicts the outcome of a given task. The second agent (agent 2) has an internal model that predicts a step-by-step state transition of the environment. By employing a rake-use task, we demonstrate that agent 2 can adapt to a task change more quickly than agent 1. The results suggest a possibility that the internal model available to each primate species can be inferred by investigating the response of the species to adapt to task changes.

The measurement of magnetoencephalographic (MEG) signals is contaminated by large magnetic artifacts, such as heart beats, eye movements, and muscle activities, and so on. These artifacts can be orders of magnitude larger than the signal from the brain, thus making cortical current estimation extremely difficult. This paper proposes a novel method to remove the effects of artifacts by simultaneously estimating the cortical and artifactual dipole currents. By using proper prior information, we show that this method can estimate the currents of artifacts and cortical activities simultaneously, and the estimated cortical currents are more reasonable in comparison to those of previous methods.

It has been suggested that the cooperation between spatial and temporal selectivities of hippocampal CA3 contributes to sequence disambiguation, which is one of essential functions of the hippocampus. We demonstrate the validity of the suggestion by using sequences assumed from a modified T-maze in which rats perform a task. It should be noted here that CA3c, which is one of the CA3 subregions, generates context-like information that contributes to sequence disambiguation. Finally, we discuss the significance of the context-like information generated in CA3c from a structural and functional viewpoint and suggest that the feedforward and feedback of the context-like information play an essential role in the memory formation of the hippocampus by affecting spatial and temporal information processing in hippocampal dentate gyrus and CA1, respectively.

Joint recognition of bio-signals emanated from human(s) is discussed. The bio-signals in this paper include camera-captured gestures and brain signals of hemoglobin change Δ

O

2

H

b

. The recognition of the integrated data is applied to the operation of a biped humanoid. Hidden Markov Models (HMMs) and Support Vector Machines (SVMs) undertake the first stage recognition of individual signal. These subsystems are regarded as soft command issuers. Then, such low-level commands are integrated by a Bayesian Network (BN). Therefore, the total system is a novel HMM/SVM-embedded BN. Using this new recognition system, human operators can control the biped humanoid through the network by realizing more motion classes than methods of HMM-alone, SVM-alone and BN-alone.

Phasic activities of dopamine (DA) neurons in the primate midbrain have been considered as representing temporal difference (TD) errors from a computational perspective. Recently, several studies have reported that, in stochastic reward tasks, the DA activities gradually increase before receiving actual rewards, which is not well explained by the simple TD model. In this study, we propose an alternative model based on a probabilistic formulation of the stochastic reward task. In simulation experiments, expectation errors, defined by the probabilistic modeling, well described the gradually increasing DA activities during a wait period even in a single trial.

A microprobing system which has the functions of measuring the intracranial EEG(IC-EEG)/temperature, providing the brain stimulation current, and freezing brain tissue is proposed for the minimally invasive brain microsurgery of intractable epilepsy treatment.

The development of cognition in humans is tightly coupled with that of language and conceptualization. It’s known that children as early as one year after birth are able to recognize the various objects as belonging to a single category. Here we report a series of experiments investigating how humans perceive the environment in terms of categorization based on the "prototype effect", studied in the literature as a characteristic feature of categorization. We focused on two parameters ("typicality" and "similarity") in the process of perceptual categorization. Using a direct rating paradigm in categorical judgment, we derived continuous parameter measures for the "typicality" and "similarity" of objects based on the subject’s estimation. The results suggest some cognitive constraints on the development of language, as it relates to the perception of sensory and motor information in the interaction with the environment.

The integration of different kinds of information is thought to be a key function of the hippocampus. On the basis of a computational theory, we hypothesize that the hippocampus performs information integration using selective desensitization of CA3 neurons, and construct a model of the hippocampal trisynaptic network on the basis of this hypothesis. This model can reproduce the results of a physiological experiment in which rat hippocampal place cells in various environments were recorded, whereas the conventional layered model cannot. This result, together with some other physiological evidence, supports our hypothesis.

Brain-computer interface (BCI) is providing a new channel for human to interact with computers and devices. Researches have been conducted on motor imaginary detection to aid individuals who cannot use any motor system for communication, to operate the computer and other devices by prosthesis. In this paper, we developed a classifier that can distinguish EEG signals responding to visual stimuli. Brightness difference was used as a preliminary case study of this approach as it is easily controllable and provides stable stimuli. The brain activities were measured by an electroencephalogram (EEG) system, and an adaptive auto-regressive (AAR) model was utilized to extract the features of different brain states corresponding to stimuli in different brightness. A minimum distance analysis (MDA) classifier was created to discriminate different brightness perception. By this means, different perceptions of four brightness conditions were decoded successfully based on the brain activities of single trials.

The usually ‘accepted’ limits of Steady State Visual Evoked Potentials are in the 3-60 Hz range. Recent studies reported SSVEP activities below 3 Hz, which remains a matter of debate. We recorded EEG responses to stimuli from 0.5 to 13 Hz. We first confirm the possibility to elicit SSVEP below 3 Hz. Afterwards, for the first time, we show that SSVEP recorded in the

δ

(0.5-5 Hz) range seem to describe several subsystems, with peaks near 1, 2.5 and 5 Hz (close to subharmonics of 10 Hz). Finally, we report surprising results in the lower frequency ranges, with responses for almost all harmonics (

e.g.

15 peak responses between 0.5 and 14.5 Hz for stimuli at 1 hz).

The concept of evolution implies that fitness traits of an organism tend toward some constrained optimality. Here, the fitness trait we consider is the distribution of photoreceptors on an organism’s retina. We postulate that an organism’s photoreceptor distribution optimizes some balance between two quantities, a benefit and a cost. The benefit is defined as the area of the field of vision. The cost is defined as the amount of time spent saccading to some target in the visual field; during this time we assume nothing is seen. Three constraints are identified. First, we assume proportional noise exists in the motor command. Second, we assume saccades are a noisy process. Third, we constrain the number of total photoreceptors. This simplified model fails to predict the human retinal photoreceptor distribution in full detail. Encouragingly, the photoreceptor distribution it predicts gets us closer to that goal. We discuss possible reasons for its current failure, and we suggest future research directions.

In this study, the plastic change in the auditory cortex induced by fear conditioning with pairing of sound (Conditioned Stimulus, CS) and electric foot-shock (Unconditioned Stimulus, US) was investigated by using of an optical recording. To investigate the effect of associated learning, optical signals in the auditory cortex to CS (12 kHz pure tone) and non-CS (8 kHz and 16 kHz pure tone) were recorded before and after the conditioning. As a result, the response area only to CS increased after the conditioning. On the other hand, to investigate whether auditory information could be retrieved by electric foot-shock after associated learning or not, auditory response to foot-shock alone was also investigated. As a result, the optical response in the auditory cortex to electric foot-shock alone could not be observed before the conditioning but clearly appeared after the conditioning.

In rat CA1 pyramidal neurons under sub- and supra-threshold conditions, our previous study showed the potentials of Cantor coding, which is theoretically proposed by Tsuda and Kuroda. However, the coding could be explained by mean rate coding simply depended on input pattern history of the peak amplitude of each input. In order to confirm that Cantor coding includes any other factors except mean rate coding, we applied three spatially different patterns of electrical stimulations with similar peak amplitudes. Although the responses did not show the statistical significant self-similar property, the membrane responses show significant clustering property. Our results suggest that some factors, which does not simply depend on mean firing rate coding, was included in Cantor coding processing.

The architecture of the human cortex is uniform and hierarchical in nature. In this paper, we build upon works on hierarchical classification systems that model the cortex to develop a neural network representation for a hierarchical spatio-temporal memory (HST-M) system. The system implements spatial and temporal processing using neural network architectures. We have tested the algorithms developed against both the MLP and the Hierarchical Temporal Memory algorithms. Our results show definite improvement over MLP and are comparable to the performance of HTM.

The study of the synchronization of EEG signals can help us to understand the underlying cognitive processes and detect the learning deficiencies. The cognitive and information processing take place in different brain regions at different time. To investigate how these distributed brain regions are linked together and the information is exchanged with time, this paper proposes a modern time-frequency coherent analysis that employs an alternative way for quantifying synchronization with both temporal and spectral resolution. Wavelet coherent spectrum is defined and computed from the EEG data set such that the cross wavelet magnitude spectra serves to indicate the degree of coherence and the cross wavelet phase can be used to provide the direction of information flow between channels on different brain regions. Several real ERP data are collected based on the cognitive tasks of sentences identification in both English and Chinese. It is observed from the time-varying synchronization that there are obviously differences during identifying both Chinese and English sentences.

Understanding the neural mechanism of object recognition is one of the fundamental challenges of visual neuroscience. However, little is known about how the information about a whole object and its parts are represented in inferior temporal (IT) cortex. To address this issue, we focus on the neural mechanism of face perception. To investigate the mechanism, we made a model of IT cortex, which performs face perception via an interaction between different IT networks. The model was made based on the face information processed by three resolution maps in early visual areas. The network model of IT consists of four kinds of networks, in which the information about a whole face is combined with the information about its face parts and their arrangements. We show here that the learning of face stimuli makes the functional connection between these IT networks, causing synchronous firing of IT neurons. The model seems to be compatible with experimental data about dynamic properties of IT neurons.

The notion of information separation in electrophysiological recordings is discussed. Whereas this problem is not entirely new, a novel approach to separate muscular interference from brain electrical activity observed in form of EEG is presented. The EEG carries brain activity in form of neurophysiological components which are usually embedded in much higher in power electrical muscle activity components (EMG, EOG, etc.). A novel multichannel EEG analysis approach is proposed in order to discover representative components related to muscular activity which are not related to ongoing brain activity but carry common patterns resulting from non-brain related sources. The proposed adaptive decomposition approach is also able to separate signals occupying same frequency bands what is usually not possible with contemporary methods.

Consensus clustering is a well studied methodology to find partitions through the combination of different formulations or clustering partitions. Different approaches for dealing with this issue using graph clustering have been proposed. Additionally, strategies to find consensus partitions by using graph-based ensemble algorithms have attracted a good deal of attention lately. A particular class of graph clustering algorithms based on spectral theory, named spectral clustering algorithms, has been successfully used in several clustering applications. However, in spite of this, few ensemble approaches based on spectral theory has been investigated. This paper proposes a consensus clustering algorithm based on spectral clustering. Preliminary results presented in this paper show the good potential of the proposed approach.

This paper investigates the dynamics of cell migration, which is the movement of a cell towards a certain target area. More specifically, the objective is to analyze the causal interdependence between cellular-morphological events and molecular-signaling events. To this end, a novel data analysis method is developed: first the local morphological changes and molecular signaling events are determined by means of edge evolution tracking (EET), next the interdependence of those events is quantified through the method of stochastic event synchrony (SES).

The proposed method is applied to time-lapse fluorescence resonance energy transfer (FRET) images of Rac1 activity in motile HT1080 cells; the protein Rac1 is well known to induce filamentous structures that enable cells to migrate. Results show a significant delay between local Rac1 activity events and morphological events. This observation provides new insights into the dynamic relationship between cellular-morphological change and molecular-signaling of migrating cells, and may pave the way to novel biophysical models of cell migration.

To detect or discover motifs in DNA sequences, two important concepts related to existing computational approaches are motif model and similarity score. One of motif models, represented by a position frequency matrix (PFM), has been widely employed to search for putative motifs. Detection and discovery of motifs can be done by comparing kmers with a motif model, or clustering kmers according to some criteria. In the past, information content based similarity scores have been widely used in searching tools. In this paper, we present a mismatch-based matrix similarity score (namely, MISCORE) for motif searching and discovering purpose. The proposed MISCORE can be biologically interpreted as an evolutionary metric for predicting a kmer as a motif member or not. Weighting factors, which are meaningful for biological data mining practice, are introduced in the MISCORE. The effectiveness of the MISCORE is investigated through exploring its separability, recognizability and robustness. Three well-known information content-based matrix similarity scores are compared, and results show that our MISCORE works well.

Due to the imprecise nature of biological experiments, biological data are often characterized by the presence of redundant and noisy data, which are usually derived from errors associated with data collection, such as contaminations in laboratorial samples. Gene expression data represent an example of noisy biological data that suffer from this problem. Machine Learning algorithms have been successfully used in gene expression analysis. Although many Machine Learning algorithms can deal with noise, detecting and removing noisy instances from data can help the induction of the target hypothesis. This paper evaluates the use of distance-based pre-processing techniques in gene expression data, analyzing the effectiveness of these techniques and combinations of them in removing noisy data, measured by the accuracy obtained by different Machine Learning classifiers over the pre-processed data. The results obtained indicate that the pre-processing techniques employed were effective for noise detection.

In recent years, the number of individuals with disabled motor functions has increased. However, restoring motor functions is possible by providing electrical stimulation to the peripheral nerves. Flexible operation methods are desirable for .people with central nerve disorders as they can be used in daily life despite limited residual motor functions. Various examples using electromyogram (EMG), joint movement or breathing to control functional electrical stimulation (FES) have been reported. The present work investigates eye direction recognition experiments and the possibility of controlling the angle of the forearm using electrical stimulation. The target angle for the forearm is determined using electrooculogram (EOG) to allow more natural operation. In addition, PD controller and disturbance observer are applied to realize forearm movement and controlled stimulation.

Detection of subtle lesions on computed tomography (CT) images is a difficult task for radiologists, because subtle lesions such as small lung nodules tend to be low in contrast, and a large number of CT images must be interpreted in a limited time. One of the solutions to the problem, a temporal subtraction technique, has been introduced in the medical field as 2D visual screening. The temporal subtraction image is obtained by subtraction of a previous image from a current one, and can be used for enhancing interval changes on medical images by removing most of the normal background structures. In this study, we have developed a method for computerized detection of lung nodules by using temporal subtraction with a voxel-matching technique in multidetector-row CT (MDCT) images. First, the candidates for nodules were detected by use of a multiple threshold technique based on the pixel value in the temporal subtraction image obtained by the voxel-matching technique. Next, a number of features were quantified, and some false positives were removed by a rule-based method with an artificial neural network. We applied our computerized scheme to 6 chest MDCT cases including 94 lung nodules. Our scheme for detecting lung nodules provided a sensitivity of 71.2 % for lung nodules with sizes under 20 mm, with 9.8 and 11.5 false positives per scan on the consistency test and validation test, respectively.

Mass spectrometry (MS) is a key technique for the analysis and identification of proteins. A prediction of spectrum peak intensities from pre computed molecular features would pave the way to a better understanding of spectrometry data and improved spectrum evaluation. The goal is to model the relationship between peptides and peptide peak heights in MALDI-TOF mass spectra, only using the peptide’s sequence information and the chemical properties. To cope with this high dimensional data, we propose a regression based combination of feature weightings and a linear predictor to focus on relevant features. This offers simpler models, scalability, and better generalization. We show that the overall performance utilizing the estimation of feature relevance and re-training compared to using the entire feature space can be improved.

Recognition of transcription factor binding sites (TFBSs or DNA motifs) to help with understanding the regulation of gene expression is one of the major challenges in the post-genomics era. Computational approaches have been developed to perform binding sites discovery based on brute-force search techniques or heuristic search algorithms, and numbers of them have achieved some degrees of success. However, the prediction accuracy of the algorithm can be relatively influenced by the natural low signal-to-noise ratio of the DNA sequence. In this paper, a novel DNA motif discovery approach using a genetic algorithm is proposed to explore the ways to improve the algorithm performance. We take account of the publicly available motif models such as Position Frequency Matrix (PFM) to initialize the population. By considering both conservation and complexity of the DNA motifs, a novel fitness function is developed to better evaluate the motif models during the evolution process. A final model refinement process is also introduced for optimizing the motif models. The experimental results demonstrate a comparable (superior) performance of our approach to recently proposed two genetic algorithm motif discovery approaches.

We propose a hybrid classification system for predicting peptide binding to major histocompatibility complex (MHC) molecules. This system combines Support Vector Machine (SVM) and Stabilized Matrix Method (SMM). Its performance was assessed using ROC analysis, and compared with the individual component methods using statistical tests. The preliminary test on four HLA alleles provided encouraging evidence for the hybrid model. The datasets used for the experiments are publicly accessible and have been benchmarked by other researchers.

In this paper, we present the performance of machine learning-based methods for detection of phishing sites. We employ 9 machine learning techniques including AdaBoost, Bagging, Support Vector Machines, Classification and Regression Trees, Logistic Regression, Random Forests, Neural Networks, Naive Bayes, and Bayesian Additive Regression Trees. We let these machine learning techniques combine heuristics, and also let machine learning-based detection methods distinguish phishing sites from others. We analyze our dataset, which is composed of 1,500 phishing sites and 1,500 legitimate sites, classify them using the machine learning-based detection methods, and measure the performance. In our evaluation, we used

f

1

measure, error rate, and Area Under the ROC Curve (AUC) as performance metrics along with our requirements for detection methods. The highest

f

1

measure is 0.8581, the lowest error rate is 14.15%, and the highest AUC is 0.9342, all of which are observed in the case of AdaBoost. We also observe that 7 out of 9 machine learning-based detection methods outperform the traditional detection method.

In this paper, we try to develop a machine learning-based virus email detection method. The key feature of this paper is employing Mail Header and Encoding Anomaly(MHEA) [1]. MHEA is capable to distinguish virus emails from normal emails, and is composed of only 5 variables, which are obtained from particular email header fields. Generating signature from MHEA is easier than generating signature by analyzing a virus code, therefore, we feature MHEA as signature to distinguish virus emails. At first, we refine the element of MHEA by association analysis with our email dataset which is composed of 4,130 virus emails and 2,508 normal emails. The results indicate that the one element of MHEA should not be used to generate MHEA. Next, we explore a way to apply MHEA into detection methods against virus emails. Our proposed method is a hybrid of matching signature from MHEA(signature-based detection) and detecting with AdaBoost (anomaly detection). Our preliminary evaluation shows that

f

1

measure is 0.9928 and error rate is 0.75% in the case of our hybrid method, which outperforms other types of detection methods.

Polymorphic viral code with encrypted payload and obfuscated decipher routine is hard to detect by generic signature scan. In this paper we propose a faster parameter detection of polymorphic viral code using hot list strategy. Parameter detection is formulated as solving SAT problem using resolution and substitution by FoL (First order Logic) theorem prover. To make parameter detection faster, we discuss one of ATP (Automated Theorem Proving) strategies, called hot list. Experiment shows that with proper selection of hot list, we can make reasoning process faster with reduction rate of generated clauses from 60% to 80%.

Coupled with the explosion of number of the network-oriented applications, intrusion detection as an increasingly popular area is attracting more and more research efforts, especially in anomaly intrusion detection area. Literature shows clustering techniques, like

K

-means, are very useful methods for the intrusion detection but suffer several major shortcomings, for example the value of

K

of

K

-means is particularly unknown, which has great influence on detection ability. In this paper, a heuristic clustering algorithm called G-means is presented for intrusion detection, which is based on density-based clustering and

K

-means and overcomes the shortcomings of

K

-means. The results of experiments show that G-means is an effective method for the intrusion detection with the high Detection Rate and the low False Positive Rate, as it can reveal the number of clusters in the dataset and initialize reasonably the cluster centroids, which makes G-means accelerate the convergence and obtain preferable performance than

K

-means.

In network environment, time-varying traffic patterns make the detection model not characterize the current traffic accurately. At the same time, the deficiency of training samples also degrades the detection accuracy. This paper proposes an anomaly detection algorithm for evolving data stream based on semi-supervised learning. The algorithm uses data stream model with attenuation to solve the problem of the change of traffic patterns, as while as extended labeled dataset generated from semi-supervised learning is used to train detection model. The experimental results manifest that the algorithm have better accuracy than those based on all historical data equivalently by forgetting historical data gracefully, as while as be suitable for the situation of deficiency of labeled data.

Malwares are spread all over cyberspace and often lead to serious security incidents. To grasp the present trends of malware activities, there are a number of ongoing network monitoring projects that collect large amount of data such as network traffic and IDS logs. These data need to be analyzed in depth since they potentially contain critical symptoms, such as an outbreak of new malware, a stealthy activity of botnet and a new type of attack on unknown vulnerability, etc. We have been developing the Network Incident analysis Center for Tactical Emergency Response (NICTER), which monitors a wide range of networks in real-time. The NICTER deploys several analysis engines taking advantage of data mining techniques in order to analyze the monitored traffics. This paper describes a brief overview of the NICTER, and its data mining based analysis engines, such as Change Point Detector (CPD), Self-Organizing Map analyzer (SOM analyzer) and Incident Forecast engine (IF).

This paper presents research leading to the development of a vision-based gesture recognition system. The system comprises of three abstract layers each with their own specific type and requirements of data. The first layer is the skin detection layer. This component provides a set of disperse skin pixels for a tracker that forms the second layer. The second component is based on the Mean-shift algorithm which has been improved for robustness against noise using our novel fuzzy-based edge estimation method making the tracker suitable for real world applications. The third component is the gesture recognition layer which is based on a gesture modeling technique and artificial neural-networks for classification of the gesture.

The strength of message authentication, digital signature and pseudonym generation mechanisms relies on the quality of the one-way hash functions used. In this paper, we propose two tests based on computational intelligence and evolutionary algorithms theory to assess the hash function quality, which may be used along with other known methods and thus comprise a testing methodology. Based on the known nonlinearity test, which might confirm uniformity of digests, we formulate two tests using Support Vector Machines (SVM)/ MLP neural networks as well as Genetic Algorithms (GA). Both tests attempt to confirm that the produced digests cannot be modeled and, moreover, that it is impossible to find two or more messages that lead to a given digest apart from involving brute force computations. Both tests are applied to confirm the quality of the well-known MD5 and SHA message digest algorithms.

The maximum clique problem, into which many problems have been mapped effectively, is of great importance in graph theory. A natural extension to this problem, emerging very recently in many real-life networks, is its fuzzification. The problem of finding the maximum clique in a fuzzy graph has been addressed in this paper. It has been shown here, that this problem reduces to an unconstrained quadratic 0-1 programming problem. Using a maximum neural network, along with, chaotic mutation capability of genetic algorithms, the reduced problem has been solved. Empirical studies have been done by applying the method on a gene co-expression network and on some benchmark graphs.

Low dimensional representation of multivariate data using unsupervised feature extraction is combined with a hybrid feature selection method to improve classification performance of recognition tasks. The proposed hybrid feature selector is applied to the union of feature subspaces selected by Fisher criterion and feature-class mutual information (MI). It scores each feature as a linear weighted sum of its interclass MI and Fisher criterion score. Proposed method efficiently selects features with higher class discrimination in comparison to feature-class MI, Fisher criterion or unsupervised selection using variance; thus, resulting in much improved recognition performance. In addition, the paper also highlights the use of MI between a feature and class as a computationally economical and optimal feature selector for statistically independent features.

Sensibility-aware image retrieval methods are presented and their performances are compared. Three systems are discussed in this paper: PCA/ICA-based method called RIM (Retrieval-aware IMage format), JPEG, and JPEG2000. In each case, a query is an image per se. Similar images are retrieved to this query. The RIM method is judged to be the best settlement in view of the retrieval performance and the response speed according a carefully designed set of opinion tests. An integrated retrieval system for image collections from the network and databases which contain RIM, JPEG and JPEG2000 is realized and evaluated lastly. Source codes of the RIM method is opened.

A recent variant of multi-task learning uses the other tasks to help in learning a task-of-interest, for which there is too little training data. The task can be classification, prediction, or density estimation. The problem is that only some of the data of the other tasks are relevant or representative for the task-of-interest. It has been experimentally demonstrated that a generative model works well in this

relevant subtask learning

task. In this paper we analyze the generalization error of the model, to show that it is smaller than in standard alternatives, and to point out connections to semi-supervised learning, multi-task learning, and active learning or covariate shift.

This paper describes the intelligent automated guided vehicle (AGV) control system. The AGV used in this paper is a virtual vehicle simulated using computer. The purpose of the control system is to control the simulated AGV for moving along the given path towards the goal. Some obstacles can be placed on or near the path to increase the difficulties of the control system. The intelligent AGV should trace the path by avoiding these obstacles. In some situations, it is inevitable to avoid the obstacles without reversing. In this paper, we look into the use of fuzzy automaton for controlling the AGV. In order to better avoid the obstacles, reverse strategy has been implemented to the fuzzy automaton controller. Another alternative to incorporate the human expertise and observations is to use a hybrid intelligent controller using fuzzy and case base reasoning to implement the reverse strategy. This paper presents the comparison results for the three intelligent AGV systems used to avoid obstacles.

This paper presents a neural network (NN) approach to determining the optimal form design of personal digital assistants (PDAs) that best matches a given set of product images perceived by consumers. 32 representative PDAs and 9 design form elements of PDAs are identified as samples in an experimental study to illustrate how the approach works. Four NN models are built with different hidden neurons in order to examine how a particular combination of PDA form elements matches the desirable product images. The performance evaluation result shows that the number of hidden neurons has no significant effect on the predictive ability of the four NN models. The NN models can be used to construct a form design database for supporting form design decisions in a new PDA product development process.

Bayesian estimation methods are used for estimation of an event rate (firing rate) from a series of event (spike) times. Generally, however, the computation of the Bayesian posterior distribution involves an analytically intractable integration. An event rate is defined in a very high dimensional space, which makes it computationally demanding to obtain the Bayesian posterior distribution of the rate.

We consider the estimation of the firing rate underlying behind a sequence that represents the counts of spikes. We derive an approximate Bayesian inference algorithm for it, which enables the analytical calculation of the posterior distribution. We also provide a method to estimate the prior hyperparameter which determines the smoothness of the estimated firing rate.

In statistical shape analysis, curve matching is often used to find correspondences between the sample points of a curve image and those of another curve image by using a dissimilarity measure of curve images. In this paper, we present a novel dissimilarity measure of curve images to be used in curve matching, together with a way of distributing sample points on each curve image. We prove that the dissimilarity measure has an asymptotic guarantee for finding a part of a curve image which is similar to a part of another one, with their respective sample points.

Recurrent neural networks are frequently used in cognitive science community for modeling linguistic structures. More or less intensive training process is usually performed but several works showed that untrained recurrent networks initialized with small weights can be also successfully used for this type of tasks. In this work we demonstrate that the state space organization of untrained recurrent neural network can be significantly improved by choosing appropriate input representations. We experimentally support this notion on several linguistic time series.

An online multibody factorization method for recovering the shape of each object from a sequence of monocular images is proposed. We formulate multibody factorization problem of data matrix of feature positions as the parameter estimation of the mixtures of probabilistic principal component analysis (MPPCA) and use the variational inference method as an estimation algorithm that concurrently performs classification of each feature points and the three-dimensional structures of each object. We also apply the online variational inference method make the algorithm suitable for real-time applications.

A number of learning machines used in information science are not regular, but rather singular, because they are non-identifiable and their Fisher information matrices are singular. Even for singular learning machines, the learning theory was developed for the case in which training samples are independent. However, if training samples have time-dependency, then learning theory is not yet established. In the present paper, we define an ergodic generalization error for a time-dependent sequence and study its behavior experimentally in hidden Markov models. The ergodic generalization error is clarified to be inversely proportional to the number of training samples, but the learning coefficient depends strongly on time-dependency.

The exchange Monte Carlo (EMC) method was proposed as an improved algorithm of Markov chain Monte Carlo method, and its effectiveness has been shown in spin-glass simulation, Bayesian learning and many other applications. In this paper, we propose a new algorithm of EMC method with Gibbs sampler by using the hidden variable representing the component from which the datum is generated, and show its effectiveness by the simulation of Bayesian learning of normal mixture models.

In this paper, we proposed a bottom-up computational model of visual filling-in to recover not only the texture but also the structure pattern in the unknown area of the images. Different from previous works of image inpainting and texture synthesis, our approach in the first step recovers the structure information of the missing part of an image; and then in the second step, each missing region with homogeneous composition is recovered independently. The structure recovery strategy is based on Gestalt laws of human visual perception, especially the good continuation law that predict the curvilinear continuity in contour completion of human behavior. In the experiment section, we provide the comparative results of our model and other proposed methods. Our model can achieve better performance in recovering images, especially when the scene contains rich structural information.

Hidden Markov Models (HMMs) are very popular generative models for sequence data. Recent research has, however, shown that Conditional Random Fields (CRFs), a type of discriminative model, outperform HMMs in many tasks. We have previously proposed Hierarchical Hidden Conditional Random Fields (HHCRFs), a discriminative model corresponding to hierarchical HMMs (HHMMs). Given observations, HHCRFs model the conditional probability of the states at the upper levels. States at the lower levels are hidden and marginalized in the model definition. In addition, we have developed a parameter learning algorithm that requires only the states at the upper levels in the training data. Previously we applied HHCRFs to the segmentation of electroencephalographic (EEG) data for a Brain-Computer Interface, and showed that HHCRFs outperform HHMMs. In this paper, we apply HHCRFs to labeling artificial data and sports video segmentation.

We apply manifold learning to a real data set of distressed and healthy companies for proper geometric tunning of similarity data points and visualization. While Isomap algorithm is often used in unsupervised learning our approach combines this algorithm with information of class labels for bankruptcy prediction. We compare prediction results with classifiers such as Support Vector Machines (SVM), Relevance Vector Machines (RVM) and the simple

k

-Nearest Neighbor (KNN) in the same data set and we show comparable accuracy of the proposed approach.

An information geometrical method is developed for characterizing or classifying neurons in cortical areas, whose spike rates fluctuate in time. When the interspike intervals of a spike sequence of a neuron obey a gamma process with a time-variant spike rate and a fixed shape parameter, the information geometry for semiparametric estimation has given the optimal method from the statistical viewpoint. Recently a more suitable statistical model for interspike intervals is proposed, which have an absolute refractory period. This work extends the information geometrical method and derives the optimal method for the new model.

This paper proposes a new method for convolutive blind speech separation by decorrelation. First, by the design of a FIR separating filter and a whitening process of the observed data, the separation of convolutive speech mixtures are transformed to find a semi-unitary separating matrix. Then we estimate the separating semi-unitary matrix by the semi-unitary joint diagonalization for a set of correlation matrices. And a numerical algortihm for semi-unitary joint diagonalziation is proposed. Simulation results of speech separation demonstrate the effectiveness of the new approach.

This paper proposes a cognitive representation and Bayesian model for spatial relations among objects that can be constructed with perception data acquired by a single consumer-grade camera. We first suggest a cognitive representation to be shared by humans and robots consisting of perceived objects and their spatial relations. We then develop Bayesian models to support our cognitive representation with which the location of a robot can be estimated sufficiently well to allow the robot to navigate in an indoor environment. Based on extensive localization experiments in an indoor environment, we show that our cognitive representation is valid in the sense that the localization accuracy improves whenever new objects and their spatial relations are detected and instantiated.

For the development of human-like intelligent robots, we have asserted the significance to introduce a general and autonomous learning system in which one neural network simply connects from sensors to actuators, and which is trained by reinforcement learning. However, it has not been believed yet that such a simple learning system actually works in the real world. In this paper, we show that without giving any prior knowledge about image processing or task, a robot could learn to approach and kiss another robot appropriately from the inputs of 6240 color visual signals in a real-world-like environment where light conditions, backgrounds, and the orientations of and distances to the target robot varied. Hidden representations that seem useful to detect the target were found. We position this work as the first step towards taking applications of the simple learning system away from “toy problems”.

To develop truly autonomous mobile robots, we proposed to introduce internal rewards such as the desire for existence, specific curiosity, diversive curiosity, boredom, and novelty into reinforcement learning. They are expected to make mobile robots capable of behaving appropriately without being told what to do. Firstly, we proposed to use multiple sources of rewards to endow mobile robots with ability to behave properly in the real world. Secondly, we proposed task-independent internal rewards. Thirdly, we proposed to attain engineering merit of internal rewards in addition to scientific interest. A pursuit-evasion game comprising a predator and its prey on a robotic field was selected as a testbed to demonstrate the effectiveness of internal rewards in reinforcement learning. The present paper focuses on learning of pursuit timing to maximize accumulated future rewards by Q-learning and SARSA.

A wheeled mobile mechanism with a passive and/or active linkage mechanism for travel in rough terrain is developed and evaluated. In our previous research, we developed a switching controller system for wheeled mobile robots in outdoor environment. This system consists of two sub-systems: an environment recognition system using a self-organizing map and an adjusted control system using a neural network. In this paper, we propose a new controller design method based on a neural network. The proposed method involves three kinds of controllers: an elementary controller, adjusted controllers, and simplified controllers. In the experiments, our proposed method results in less oscillatory motion in outdoor environment and performs better than a well tuned PID controller does.

Sensing the distance to an object is a very important clue in many fields and is an area of active research. In our research we propose an alternative method to tell the distance by using motion parallax method that is based on monocular vision and a moving observer with a constant acceleration approximately in an unknown environment. In this paper, our proposed method is applied for the outdoor environment and showed well performance. This is a good point for mobile robots or vehicles to obtain depth by using this method as long as the observer acceleration is known.

The attitude control of a horizontal underactuated Manipulator (UAM) is a difficult control problem because of its second-order nonholonomic constraints. This paper proposes a new method for the trajectory planning of the UAM. In the proposed method, the trajectory planning is transformed to a constraint satisfaction problem, and the constraint satisfaction problem is solved by using a sequence of multi-layer perceptrons which are trained to be a forward model of the UAM. We show the effectiveness of our method by several experimental results.

This paper shows that neural-net based machines may be designed to mimic the consciousness-sensations generated by the human brain. It is shown that the standard definition of biological modalities of the tactile and visual receptors, coupled with the law of specific nerve energy, leads to a fundamental relationship that relates human subjective experiences, or consciousness, to explicit neuronal activity. Such a relationship is a giant leap forward in the study of consciousness since it converts the parameters of consciousness, which have never been amenable to mathematical calculations, into mathematically calculable functions.

We present a category learning vector quantization (cLVQ) approach for incremental and life-long learning of multiple visual categories where we focus on approaching the stability-plasticity dilemma. To achieve the life-long learning ability an incremental learning vector quantization approach is combined with a category-specific feature selection method in a novel way to allow several metrical “views” on the representation space for the same cLVQ nodes.

We present a biologically inspired vision system able to incrementally learn multiple visual categories by interactively presenting several hand-held objects. The overall system is composed of a foreground-background separation part, several feature extraction methods and a life-long learning approach combining incremental learning with category specific feature selection. In contrast to most visual categorization approaches where typically each view is assigned to a single category we allow labeling with an arbitrary number of shape and color categories and also impose no restrictions to the viewing angle of presented objects.

In this paper, we propose a new multitask learning (MTL) model which can learn a series of multi-class pattern recognition prob- lems stably. The knowledge transfer in the proposed MTL model is implemented by the following mechanisms: (1) transfer by sharing the internal representation of RBFs and (2) transfer of the information on class subregions from the related tasks. The proposed model can detect task changes on its own based on the output errors even though no task information is given by the environment. It also learn training samples of different tasks that are given one after another. In the experiments, the recognition performance is evaluated for the eight MTPR problems which are defined from the four UCI data sets. The experimental results demonstrate that the proposed MTL model outperforms a single-task learning model in terms of the final classification accuracy. Furthermore, we show that the transfer of class subregion contributes to enhancing the generalization performance of a new task with less training samples.

The hierarchical structure of real-world problems has resulted in a focus on hierarchical frameworks in the reinforcement learning paradigm. Preparing mechanisms for automatic discovery of macro-actions has mainly concentrated on subgoal discovery methods. Among the proposed algorithms, those based on graph partitioning have achieved precise results. However, few methods have been shown to be successful both in performance and also efficiency in terms of time complexity of the algorithm. In this paper, we present a SCC-based subgoal discovery algorithm; a graph theoretic approach for automatic detection of subgoals in linear time. Meanwhile a parameter tuning method is proposed to find the only parameter of the method.

This paper studies a piecewise linear chaotic spiking oscillator relating to neuron models. Repeating vibrate-and-fire dynamics, the system can exhibit chaotic/periodic spike-trains and related bifurcation phenomena. Deriving the return map of a state variable, we can analyze typical phenomena precisely and have confirmd an interesting bifurcation phenomena of chaotic spike-trains and window structure of period-doubling route.

This paper studies typical nonlinear dynamics of spiking circuit including two capacitors. Applying impulsive switching depending on both state and time, the circuit can exhibit rich chaotic/periodic phenomena. We pay special attention to superstable periodic orbits and related bifurcation phenomena. The circuit dynamics can be simplified into a piecewise linear one-dimensional return map that enables us to analyze basic bifurcation phenomena precisely.

We propose a distributed radio access network selection method for heterogeneous wireless network environment, in which mobile terminals can adaptively and seamlessly handover among different wireless access technologies. Our algorithm optimizes fairness of radio resource usage without centralized computing on the network side. As a decentralized optimization scheme, we introduce the dynamics of the mutually connected neural network dynamics, whose energy function autonomously minimizes by distributed update of each neuron. Since the objective function of the fairness becomes a fourth-order function of the neurons’ states which cannot be optimized by the conventional Hopfield neural network, we apply a neural network model extended to higher-order mutual connections and energy functions. By numerical simulation, we confirm that the proposed algorithm can optimize fairness of the throughput by distributed and autonomous computation.

The brain is capable of parallel processing of different types of information in different brain regions. However, for higher cognitive functions such as decision-making, such regionally distributed information must interact together in the proper timing. The prefrontal cortex, the region which carries out decision-making, needs to be modulated by external signals to represent the current behavioral context in the hippocampus (HP). The question remains as to how the firing activity of the cortical neural network can be modulated by external signals in corporation with the ongoing activity. We hypothesized that rhythmic signals that attempt to synchronize the cortical ongoing activity minimize the disturbance and effectively enhance the activities of selective neurons. We investigated the level of the modulation by using a mutually connected neural network that consists of a neuron model with excitatory and refractory periods. The results demonstrated that cortical ongoing activities are weakly modulated by random external signals, while synchronized rhythmic signals, given as the pseudo HP signals, selectively enhance cortical activities. This suggests that the cortical ongoing activity is effectively influenced by the synchronized signals, which carry information in the proper timing of excitation. The investigation of neural synchronization dynamics is important to understanding how the brain realizes parallel processing in different sub-regions and to update immediately the internal representation even if the previous internal processing is ongoing.

We study synchronization phenomena in a pair of integrate-and-fire (IF) oscillators with the width of an action potential. They have slightly different periodic firings each other. Such non-identical IF oscillators typically interact with each other via excitatory or inhibitory synaptic stimuli. A return map analysis gives us a systematic analysis of stabilities of the 1:1 phase locking states in the pair, in terms of a synaptic decaying parameter. We demonstrate transitions of the phase lockings with a small difference in their firing frequencies and different width of an action potential. Spike timing histograms are also observed in the pair with dynamical noise. The slight shifts of a synchronous state effectively induces distribution shifts in the spike timing histogram. This agrees with results observed in physiological experiments.

The Inverse function Delayed (ID) model is a novel neuron model derived from a macroscopic model which is attached to conventional network action. The special characteristic of the ID model is to have the negative resistance effect. Such a negative resistance can actively destabilize undesirable states, and we expect that the ID model can avoid the local minimum problems for solving the combinatorial optimization problem. In computer simulations, we have shown that the ID network can avoid the local minimum problem with a particular combinatorial optimization problem, and we have also shown the existence of an appropriate parameter for finding an optimal solution with high success rate experimentally. In this paper, we theoretically estimate appropriate network parameters to remove all local minimum states.

This paper proposes Fractional-order Hopfield Neural Networks (FHNN). This network is mainly based on the classic well-known Hopfield net in which fractance components with fractional order derivatives, replace capacitors. Stability of FHNN is fully investigated through energy-like function analysis. To show how effective the FHNN network is, an illustrative example for parameter estimation problem of the second-order system is finally considered in the paper. The results of simulation are very promising.

This paper describes the basic steps of a novel approach to weather classification by remote and local atmosphere sensing. Atmospheric data on the troposphere are gathered by 10 GHz radio links and a number of meteorological sensors. Classification is performed by artificial neural networks (ANN) and is crucial for further processing, because of the different RF propagation influences under a variety of weather conditions. Reasons for using ANN compared to other means of classification are discussed. Differences in the size and number of hidden layers of back-propagation networks used are discussed. Different learning sets of measured data and their construction are also evaluated.

Predictions of the distribution of fungal crop diseases have previously been made solely from climatic data. To our knowledge there has been no study that has used biotic variables, either alone or in combination with climate factors, to make broad scale predictions of the presence or absence of fungal species in particular regions. The work presented in this paper used multi-layer perceptrons (MLP) to predict the presence and absence of several species of fungal crop diseases across world-wide geographical regions. These predictions were made using three sets of variables: abiotic climate variables; biotic variables, represented by host plant assemblages; And finally the combination of predictions of the climate and host assemblage MLP using a cascaded MLP architecture, such that final predictions were made from both abiotic and biotic factors.

Environmental changes in oceanic conditions have the potential to cause jellyfish populations to rapidly expand leading to ecosystem level repercussions. To predict potential changes it is necessary to understand how such populations are influenced by oceanographic conditions. Data recording the presence or absence of jellyfish of the genus

Physalia

at beaches in the West Auckland region of New Zealand were modelled using Multi-Layer Perceptrons (MLP) with time lagged oceanographic data as input data. Results showed that MLP models were able to generalise well based on Kappa statistics and gave good predictions of the presence or absence of

Physalia

. Moreover, an analysis of the network contributions indicated an interaction between wave and wind variables at different time intervals can promote or inhibit the occurrence of

Physalia

.

The motivation for modelling the effects of climate change on viticulture and wine quality using both quantitative and qualitative data within an integrated analytical framework is described. The constraints and solutions evident when taking such an approach are outlined. WEBSOM is a novel self-organising map (SOM) method for extracting relevant domain-dependent characteristics from web based texts and in this case, investigated for modelling wine quality resulting from climate variation, by web text mining published descriptions made by sommeliers about this phenomenon. This paper describes experiments using the WEBSOM method with their results.

In order to make a support vector machine applicable to time-varying problems, a forgetting factor is introduced to its cost function, in the same way as the RLS algorithm for adaptive filters. The idea of the forgetting factor is simple but it is shown to drastically change the performance of SVMs by deriving the average generalization error in a simple case where input space is one-dimensional. The average generalization error does not converge to zero, differently from the SVM in batch or online. We confirmed our results by computer simulations.

We propose two methods for tuning membership functions of a fuzzy classifier by the support-vector-machine (SVM) like training. For each class, we define a membership function in the feature space. In the first method, we tune the slopes of the membership functions so that the margin between classes is maximized. This method is similar to a linear all-at-once SVM. We call this AAO tuning. In the second method, for each class the membership function is tuned so that the margin between the class and the remaining classes are maximized. This method is similar to a linear one-against-all SVM. This is called OAA tuning. According to the computer experiment, the kernel-discriminant-analysis (KDA) based fuzzy classifiers tuned by AAO tuning and by OAA tuning and SVM show comparable classification performance.

This paper addresses the problem of excessively long classifier training times associated with using the Adaboost algorithm within the framework of a cascade of boosted ensembles (CoBE). We present new test results confirming the acceleration of the training phase and the robustness of the

Parallel

Strong

classifier within the same

Layer

(PSL) training structure recently proposed by [1]. The findings demonstrate a speed up of an order of magnitude over the current training methods without a compromise in accuracy. We also present a modified version of the PSL training structure that further decreases the duration of the training phase while preserving accuracy.

This paper presents a method, which enables a robot to extract demonstrator’s key motions based on imitation learning through unsegmented human motion. When a robot learns another’s motions from unsegmented time series, the robot has to find what he learns from the continuous motion. The learning architecture is developed mainly based on a switching autoregressive model (SARM), a simple phrase extraction method, and singular vector decomposition to discriminate key motions. In most previous research on methods of imitation learning by autonomous robots, target motions that were given to robots were segmented into several meaningful parts by the experimenters in advance. However, to imitate certain behaviors from the continuous motion of a person, the robot needs to find segments that should be learned. In our approach, the learning architecture converts the continuous time series into a discrete time series of letters by using SARM, finds candidates of key motions by using a simple phrase extractor which utilizes n-gram statistics, and removes meaningless segments from the keywords by utilizing singular vector decomposition (SVD) to achieve this goal,. In our experiment, a demonstrator displayed several unsegmented motions to a robot. The results revealed that the framework enabled the robot to obtain several prepared key motions.

This paper presents our development on mobile robot in charge of security in an office environment. The developed robot uses curiosity on environmental change for threat detection. The platform used to develop the security robot is the ‘WITH’ mobile robot [1], imported from Kitakyushu Institute of Technology Japan. Within the environment of a research office, the WITH security robots capability has been tested on threat detection, programmable security navigation and threat tracking. The results achieved from the tests have been highlighted below: (1) Environment curiosity modelling: this allows for more dynamic threat detection than object recognition method, (2) Programmable navigation system: enables multi-points security monitoring.

The

all together

model is one of the support vector machine (SVM) for multiclass classification by using a piece-wise linear function. As a novel all together model, we already proposed a hard-margin multiobjective SVM model for piecewise linearly separable data, which maximizes all of the geometric margins simultaneously for the generalization ability. In addition, we derived a single-objective convex problem whose optimal solution is weakly Pareto optimal for the proposed SVM. However, in the real-world classification problem the data are often piecewise linearly inseparable. Therefore, in this paper we extend the hard-margin SVM for the data by introducing penalty functions for the margin slack variables based on the geometric distances between outliers and the support hyperplanes, and incorporating those functions into the objective functions. Moreover, we derive a single-objective second-order cone programming (SOCP) problem, and show that its optimal solution is weakly Pareto optimal for the proposed soft-margin SVM. Furthermore through numerical experiments we verify that the SOCP model maximizes the geometric margins in the sense of multiobjective optimization.

In order to develop a method to infer the underlying synaptic connectivity only from spatio-temporal patterns of spiking activity observed in a neuronal network, we here investigated characteristics of a network of functional connections defined by pairwise spike synchrony. We first conducted numerical simulations of a computational model of a local cortical network and constructed a functional network based on the obtained spike data. The proposed analysis with the optimal parameters defining functional connections showed that characteristics of connectivity of functional networks are in good agreement with those of synaptic connectivity in the computational model in terms of statistical indices such as the clustering coefficient and the shortest path length. The result suggests that it is possible to extract at least statistical characteristics of synaptic connectivity from spatio-temporal patterns of spiking activity.

Glycosylation is one of the main topics in understanding the life systems. More than a half of the protein is glycosylated to acquire the function, structural stability and biological diversity.

O

-glycosylation is one of the two main types of the mammalian protein glycosylation. Though it is known to serine or threonine specific, any consensus sequence is still unknown, while the binding process and the consensus sequence are clarified for the other type of

N

-glycosylation. We use support vector machines (SVM) for the prediction of

O

-glycosylation sites using the experimental data as the input information such as protein primary sequences, structural and biochemical characters around a prediction target aiming to elucidate the glycosylation mechanism and the existence of any motives. The present paper also reports the results obtained by the semi-supervised learning using transductive SVM considering a possibility of unobserved glycosylation sites, and by the marginalized kernel considering hidden variables.

For precise estimation with soft sensors, it is necessary to remove outliers from the measured raw data before constructing the model. Conventionally, visualization and maximum residual error have been used for outlier detection, but they often fail to detect outliers for nonlinear function with multidimensional input. In this paper we propose a practical approach to outlier detection using Support Vector Regression, which reduces computational cost and defines outlier threshold appropriately. We apply this approach to both test and industrial datasets for validation.

This paper proposes a special adaptive mean shift clustering algorithm, especially for the case of highly overlapping clusters. Its application is demonstrated for simulated data, aiming at finding the ‘old clusters’. The obtained clustering result is actually close to an estimated upper bound, derived for those simulated data elsewhere.

SNNs have been tested as possible candidates for the implementation of robot controllers, in particular behaviour based controllers, but in most approaches their real power, related to their inherent temporal processing, and, especially, temporal pattern generating capabilities, have been ignored. This paper is concerned with showing how SNNs in their most dynamic form can be easily evolved to provide the adaptable or sensor and context modulated pattern generating capabilities required for the generation of action sequences in robots. In fact, the objective is to have a structure that can provide a sequence of actions or a periodic pattern that extends in time from a very time limited sensorial cue.

A neuro-evolutive system for mobile robot ego-motion estimation using time-of-flight (TOF) 3D camera readings is presented in this paper. It is composed of two modules. First, a Neural Gas adaptative algorithm is used to obtain a set of codevectors quantizing the preprocessed 3D measurements provided by the camera. Second, codevector sets from consecutive frames are matched using an evolutive algorithm in order to estimate the motion of the robot between those two positions.

In this paper the fusion of artificial neural networks, granular computing and learning automata theory is proposed and we present as a final result ANLAGIS, an adaptive neuron-like network based on learning automata and granular inference systems. ANLAGIS can be applied to both pattern recognition and learning control problems. Another interesting contribution of this paper is the distinction between pre-synaptic and post-synaptic learning in artificial neural networks. To illustrate the capabilities of ANLAGIS some experiments with multi-robot systems are also presented.

This paper presents an incremental learning algorithm primarily intended for adjusting and optimizing high dimensional backpropagation ANN-based supervised classification systems. The algorithm avoids the highly time-consuming pre-processing stage used to reduce dimensionality through the deletion or averaging of redundant information and the establishment of an appropriate processing window. The proposed algorithm acts during the training process of the ANN by automatically obtaining the optimal window size and transformation parameter values needed for a given set of classification requirements. During this process, it changes the network topology on line by adaptively appending input units and their corresponding connections to the existing network. The proposed process allows the ANN to learn incrementally by adapting to the new topology without forgetting what had been learnt earlier. This process could also be used as an incremental learning system in, for instance, robotic systems; when new instances are fed, as it would not need to perform a whole training stage. Instead, the knowledge encoded in the new instances could be learnt through the proposed adjustment of network topology.

This paper reports initial steps in the study of control strategies for a multi-robot system trying to move a flexible hose. Our starting point is the hose geometry modeling using cubic splines. The control problem is then stated as the problem of reaching a desired configuration of the spline control points from an initial configuration. The control of the hose by the multi-robot system is first solved neglecting the hose internal dynamics. This is the simplest ideal case. Then we give a model for the internal dynamics. Future and on going work is addressing the definition of a control system that takes into account the hose internal dynamic constraints.

In spite of the multiple advantages that multi-robot systems offer, to turn them into a realistic option and to get their proliferation, they must be economically attractive. Multi-robot systems are composed of several robots that generally are similar, so if an economic optimization is done in one of them, such optimization can be replicated in each member. In this paper we deal with the economic optimization of each control loops of the subsystems that each robot must control individually. As the subsystems can be complex, we propose to use a Predictive Control modeled by Time Delayed Neural Networks and implemented using very low cost Field Programmable Gate Arrays.

A novel approach is presented for adaptive trajectory tracking of robot manipulators using a three-stage hierarchical neural network model involving Support Vector Machines (SVM) and an adaptive unsupervised Neural Network. It involves a novel adaptive Self Organizing feature map (SOFM) in the first stage which aims at clustering the input variable space into smaller sub-spaces representative of the input space probability distribution and preserving its original topology, while rapidly increasing, on the other hand, cluster distances. Moreover, its codebook vector adaptation rule involves m-winning neurons dynamics and not the winner takes all approach. During convergence phase of the map a group of Support Vector Machines, associated with its codebook vectors, is simultaneously trained in an online fashion so that each SVM learns to positively respond when the input data belong to the topological sub-space represented by its corresponding codebook vector, taking into account similarity with that codebook vector. Moreover, it learns to negatively respond to input data not belonging to such a previously mentioned corresponding topological sub-space. The proposed methodology is applied, with promising results, to the design of a neural-adaptive trajectory tracking controller, by involving the computer-torque approach, which combines the proposed three-stage neural network model with a classical servo PD feedback controller. The results achieved by the suggested hierarchical SVM approach are favorably compared to the ones obtained by traditional (PD) and non-hierarchical neural network based controllers.

This paper presents an application of a Fuzzy Colour Contrast Fusion (FFCF) algorithm in compensating for reduced colour depth representation of a colour image while maintaining efficient colour sensitivity that suffices for accurate real-time colour-based object recognition. We investigate the effects of applying fuzzy colour contrast rules to varying colour depth as we extract the optimal rule combination. The experiments were performed using the robot soccer game set-up with spatially varying illumination intensities on the scene. Interestingly, our results show that for most cases, colour depth reduction could actually improve colour classification via a pie-slice technique, in a modified rg-chromaticity colour space. For 6 different colours, the algorithm was able to yield 6.5% higher overall accuracy with only one-twelfth of LUT size than the full colour depth LUT.

Intelligent navigation and path-finding for computer- animated characters in graphical 3D environments is a major design challenge facing programmers of simulations, games, and cinematic productions. Designing agents for computer-animated characters that are required to both move intelligently around obstacles in the environment, and do so in a psycho-visually realistic way with smooth motion is often a too-difficult challenge - designers generally sacrifice intelligent navigation for realistic movement or vice-versa. We present here a specially adapted hybrid fuzzy A* algorithm as a viable solution to meet both of these challenges simultaneously. We discuss the application of this algorithm to animated characters and outline our proposed architecture for automatic tuning of this system.

In biological systems, internal microorganisms adaptively survive but often serve necessary functions to the host, such as energy production by mitochondria as symbiosis. On the other hand, malignant germs have a parasitic relationship with the host and provide no benefit. A significant property that distinguishes healthy symbioses and malignant parasites is reproduction speed, or pace. For example, the rapid reproduction of influenza viruses destructs the host system, resulting in death. This study explored the necessary temporal property to establish a healthy relationship with the host under conditions where internal organisms have individual life spans. We propose a simple model of microorganisms, which are distributed spatially as colonial organizations undergoing temporal evolution and hypothesize that a self-consistent rhythm generated in collective behavior that is functionally coupled with the temporal global property of the host system is critical. To investigate the real-time coordination capability, an experimental framework with a mobile robot moving in the real world was used. As the on-line system, the microorganism model controls this robot. In this model, microorganisms expanded spatially and had colonial and power law distributions through time evolution. The neighboring distances, which are crucial for reproduction speed and are globally modulated by the size of the whole living area, are plastically changed to exhibit a rhythmic modulation. In the real-environmental experiment, the robot’s navigation was successfully demonstrated by producing a temporal adaptability of microorganisms with the living area reshaped according to the current sensory information of the mobile robot. This is a first step of the microorganism-based framework to investigate the real-time coordination mechanism between internal and external timescales. The result may further groundbreaking research of bio-morphological robots.

This paper introduces a stream programming based design of the zero and higher order central moments that use an integral image or summed area data structure of geometric moments. The stream programming algorithm runs on a general purpose graphics processing unit (GPGPU) that are becoming commodity hardware, giving real-time performance even for large image sizes and a large number of scan window sizes.

In this paper, the Bayesian fusion of auditory and visual spatial cues has been implemented in a humanoid robot aiming to increase the accuracy of localization, given a situation that an audiovisual stimulus was presented. The performance of auditory and visual localization was tested under two conditions: fixation and saccade. In this experiment, we proved that saccade did greatly reduce the accuracy of auditory localization in the humanoid robot. The Bayesian model became not reliable when the results of auditory and visual localization were not reliable, particularly during saccade. During the tests, localization in two conditions (saccade onset and changing of direction of motion) has been ignored and only azimuth position has been considered.

A solution to the problem of simultaneous localization and mapping, known as the problem of SLAM, would be of inestimable value to the field of autonomous robots. One possible approach to this problem depends on the establishment of landmarks in the environment, using artificial structures or predetermined objects that limit their applicability in general tasks. This paper presents a solution to the problem of SLAM that relies on an omnidirectional vision system to create a sparse landmark map composed of natural structures recognized from the environment, used during navigation to correct odometric errors accumulated over time. Visual sensors are a natural and compact way of achieving the rich and wide characterization of the environment necessary to extract natural landmarks, and the omnidirectional vision increases the amount of information received at each instant. This solution has been tested in real navigational situations and the results show that omnidirectional vision sensors are a valid and desirable way of obtaining the information needed to solve the problem of SLAM.

This paper presents a novel and notable swarm approach to evolve an optimal set of weights and architecture of a neural network for classification in data mining. In a distributed environment the proposed approach generates randomly multiple architectures competing with each other while fine-tuning their architectural loopholes to generate an optimum model with maximum classification accuracy. Aiming at better generalization ability, we analyze the use of particle swarm optimization (PSO) to evolve an optimal architecture with high classification accuracy. Experiments performed on benchmark datasets show that the performance of the proposed approach has good classification accuracy and generalization ability. Further, a comparative performance of the proposed model with other competing models is given to show its effectiveness in terms of classification accuracy.

This research presents a Field Programmable Gate Array (FPGA) implementation of a taste recognition model. The model is based on simple integrate and fire neurons and facilitates an on-line learning. The whole system, including the hardware required to build (evolve) the network was hosted on one FPGA chip. The implementation used 45% of the logic elements, 76% of the memory, and 23% of the dedicated multiplier slices of the chip. FPGA size was sufficient for 64 neurons with up to 64 synapses each (a total of 4096 synapses). The proposed FPGA implementation was successfully applied to a classification problem of taste recognition and the FPGA implementation was at least 10 times faster when evolving the network and 74 times faster during the classification than the software simulations executed by a stand-alone PC.

The Hybrid neural Fuzzy Inference System (HyFIS) is a five layers adaptive neural fuzzy system for building and optimizing fuzzy models. In this paper, the fuzzy Yager inference scheme, which accounts for a firm and intuitive logical framework that emulates the human reasoning and decision-making mechanism, is integrated into the HyFIS network. In addition, a self-organizing gaussian Discrete Incremental Clustering (gDIC) technique is used to form the fuzzy sets in the fuzzification phase. This clustering technique is no longer limited by the need to have prior knowledge about the number of clusters needed in each input and output dimensions. The proposed self-organizing Hybrid neural Fuzzy Inference System based on Yager inference (HyFIS-Yager-gDIC) is benchmarked on two case studies to demonstrate its superiority as an effective neuro-fuzzy modelling technique.

This paper studies a parallel ant colony optimizer and its application to the traveling sales person problems. The parallel processing is based on the adaptive resonance theory map that divide the input space into subspaces. The ants are classified into two types: local ant for local search within either subspace and global ant for search of whole input space. Communication between local and global ants is a key for effective parallel processing. Applying the algorithm to basic bench marks, we can suggest that our algorithm realize fast and reasonable search.

Learning strategies under

covariate shift

have recently been widely discussed. Under covariate shift, the density of learning inputs is different from that of test inputs. In such environments, learning machines need to employ special learning strategies to acquire a greater capability to generalize through learning.

However, incremental learning methods are also for learning in non-stationary learning environments, which would represent a kind of covariate-shift. However, the relation between covariate shift environments and incremental learning environments has not been adequately discussed.

This paper focuses on the covariate shift in incremental learning environments and our re-construction of a suitable incremental learning method.

In this paper, we propose a new incremental linear discriminant analysis (ILDA) for multitask pattern recognition (MTPR) problems in which training samples of a specific recognition task are given one after another for a certain period of time and the task is switched to another related task in turn. The Pang et al.’s ILDA is extended such that a discriminant space of the current task is augmented with effective discriminant vectors that are selected from other related tasks based on the class separability. We call the selection and augmentation of such discriminant vectors

knowledge transfer of feature subspaces

. In the experiments, the proposed ILDA is evaluated for the four MTPR problems, each of which consists of three multi-class recognition tasks. The results demonstrate that the proposed ILDA outperforms the ILDA without the knowledge transfer with regard to both the class separability and recognition accuracy. From the experimental results, we confirm that the proposed knowledge transfer mechanism works well to construct effective discriminant feature spaces incrementally.

When it comes to application of computational learning techniques in practical scenarios, like for example adaptive inferential control, it is often difficult to apply the state-of-the-art techniques in a straight forward manner and usually some effort has to be dedicated to tuning either the data, in a form of data pre-processing, or the modelling techniques, in form of optimal parameter search or modification of the training algorithm. In this work we present a robust approach to on-line predictive modelling which is focusing on dealing with challenges like noisy data, data outliers and in particular drifting data which are often present in industrial data sets. The approach is based on the local learning approach, where models of limited complexity focus on partitions of the input space and on an ensemble building technique which combines the predictions of the particular local models into the final predicted value. Furthermore, the technique provides the means for on-line adaptation and can thus be deployed in a dynamic environment which is demonstrated in this work in terms of an application of the presented approach to a raw industrial data set exhibiting drifting data, outliers, missing values and measurement noise.

Recently, the formation of topographic maps has been approached via a direct optimization strategy involving the use of heuristic search techniques. In this paper, we move a step further in this line of research by devising and empirically assessing the performance of six different evolutionary algorithms (EA) towards the automatic generation of high-quality maps. Besides, we also report an analysis over the convexity profiles exhibited by different realizations of the adopted cost function in a manner as to testify its inherent search complexity. The simulation results reveal that, although the EA schemes do not distinguish so much in terms of the average quality of the maps they form, there is sometimes a significant difference of performance in terms of robustness (variance of the quality indices) and efficiency (number of iterations to converge to a good solution).

The inherent nonlinear of brushless DC motor (BLDCM) makes it hard to get a good performance by using the conventional PI controller to the speed control of BLDCM. In this paper, a radial basis function (RBF) artificial neural network (NN) based adaptive PI controller for BLDCM is developed. The RBF NN has a strong ability of adaptive, self-learning and self-organization. At the same time, the nonlinear mapping property and high parallel operation ability of NN make it suitable to be applied to perform parameter identification. In this paper, the RBF NN is employed to predict the Jacobian information and tune the gains. Compared with back propagation (BP) type NN with sigmoid activation function, the RBF NN has a more fast convergence speed and can avoid getting stuck in a local optimum. Through parameter prediction, response speed of the system can be improved. The experimental results demonstrate that a high control performance is achieved. The system responds quickly with little overshoot. The steady state error is zero. The system shows robust performance to the load torque disturbance.

We have proposed an online feature extraction method called Chunk Incremental Principal Component Analysis (Chunk IPCA) where a chunk of data is trained at a time to update an eigenspace model. In this paper, we propose an extended version of Chunk IPCA in which a proper threshold for the accumulation ratio is adaptively determined such that the highest classification accuracy is maintained for a validation data set. Whenever a new chunk of training data is given, the validation set is updated in an online fashion by using the

k

-means clustering or through the prototype selection based on the classification results. The experimental results show that the extended version of Chunk IPCA can determine a proper threshold on an ongoing basis, resulting in keeping higher classification accuracy than the original Chunk IPCA.

A novel ontology based chronic disease risk analysis system framework is described, which allows the creation of global knowledge representation (ontology) and personalized modeling for a decision support system. A computerized model focusing on organizing knowledge related to three chronic diseases and genes has been developed in an ontological representation that is able to identify interrelationships for the ontology-based personalized risk evaluation for chronic diseases. The personalized modeling is a process of model creation for a single person, based on their personal data and the information available in the ontology. A transductive neuro-fuzzy inference system with weighted data normalization is used to evaluate personalized risk for chronic disease. This approach aims to provide support for further discovery through the integration of the ontological representation to build an expert system in order to pinpoint genes of interest and relevant diet components.

The effect of frost on the successful growth and quality of crops is well understood by growers as leading potentially to total harvest failure. Studying the frost phenomenon, especially in order to predict its occurrence has been the focus of numerous research projects and investigations. Frost prone areas are of particular concern. Grape growing for wine production is a specific area of viticulture and agricultural research. This paper describes the problem, outlines a wider project that is gathering climate and atmospheric data, together with soil, and plant data in order to determine the inter-dependencies of variable values that both inform enhanced crop management practices and where possible, predict optimal growing conditions. The application of some novel data mining techniques together with the use of computational neural networks as a means to modeling and then predicting frost is the focus of the investigation described here as part of the wider project.

This paper presents a novel gene selection method based on personalized modeling. Identifying a compact set of genes from gene expression data is a critical step in bioinformatics research. Personalized modeling is a recently introduced technique for constructing clinical decision support systems. In this paper we have provided a comparative study using the proposed Personalized Modeling based Gene Selection method (PMGS) on two benchmark microarray datasets (Colon cancer and Central Nervous System cancer data). The experimental results show that our method is able to identify a small number of informative genes which can lead to reproducible and acceptable predictive performance without expensive computational cost. These genes are of importance for specific groups of people for cancer diagnosis and prognosis.

This study extends the recently proposed Evolving Spiking Neural Network (ESNN) architecture by combining it with an optimization algorithm, namely the Versatile Quantum-inspired Evolutionary Algorithm (vQEA). Following the wrapper approach, the method is used to identify relevant feature subsets and simultaneously evolve an optimal ESNN parameter setting. Applied to carefully designed benchmark data, containing irrelevant and redundant features of varying information quality, the ESNN-based feature selection procedure lead to excellent classification results and an accurate detection of relevant information in the dataset. Redundant and irrelevant features were rejected successively and in the order of the degree of information they contained.

The behaviour of multiple stock markets can be described within the framework of complex dynamic systems (CDS). Using a global model with the Kalman Filter we are able to extract the dynamic interaction network (DIN) of these markets. The model was shown to successfully capture interactions between stock markets in the long term. In this study we investigate the effectiveness of two different personalised modelling approaches to multiple stock market prediction. Preliminary results from this study show that the personalised modelling approach when applied to the rate of change of the stock market index is better able to capture recurring trends that tend to occur with stock market data.

This paper proposes an extension to the original offline version of DENFIS. The new algorithm, DyNFIS, replaces original triangular membership function with Gaussian membership function and use back-propagation to further optimizes the model. Fuzzy rules are created for each clustering centre based on the clustering outcome of evolving clustering method. For each test data, the output of DyNFIS is calculated through fuzzy inference system based on

m

-most activated fuzzy rules and these rules are updated based on back-propagation to minimize the error. DyNFIS shows improvement on multiple benchmark data and satisfactory result in NN3 forecast competition.