Neurocomputing

Published by Elsevier
Online ISSN: 0925-2312
Publications
Article
Texture boundary detection (or segmentation) is an important capability in human vision. Usually, texture segmentation is viewed as a 2D problem, as the definition of the problem itself assumes a 2D substrate. However, an interesting hypothesis emerges when we ask a question regarding the nature of textures: What are textures, and why did the ability to discriminate texture evolve or develop? A possible answer to this question is that textures naturally define physically distinct (i.e., occluded) surfaces. Hence, we can hypothesize that 2D texture segmentation may be an outgrowth of the ability to discriminate surfaces in 3D. In this paper, we conducted computational experiments with artificial neural networks to investigate the relative difficulty of learning to segment textures defined on flat 2D surfaces vs. those in 3D configurations where the boundaries are defined by occluding surfaces and their change over time due to the observer's motion. It turns out that learning is faster and more accurate in 3D, very much in line with our expectation. Furthermore, our results showed that the neural network's learned ability to segment texture in 3D transfers well into 2D texture segmentation, bolstering our initial hypothesis, and providing insights on the possible developmental origin of 2D texture segmentation function in human vision.
 
Article
A new technique to extract and evaluate physical activity patterns from image sequences captured by a wearable camera is presented in this paper. Unlike standard activity recognition schemes, the video data captured by our device do not include the wearer him/herself. The physical activity of the wearer, such as walking or exercising, is analyzed indirectly through the camera motion extracted from the acquired video frames. Two key tasks, pixel correspondence identification and motion feature extraction, are studied to recognize activity patterns. We utilize a multiscale approach to identify pixel correspondences. When compared with the existing methods such as the Good Features detector and the Speed-up Robust Feature (SURF) detector, our technique is more accurate and computationally efficient. Once the pixel correspondences are determined which define representative motion vectors, we build a set of activity pattern features based on motion statistics in each frame. Finally, the physical activity of the person wearing a camera is determined according to the global motion distribution in the video. Our algorithms are tested using different machine learning techniques such as the K-Nearest Neighbor (KNN), Naive Bayesian and Support Vector Machine (SVM). The results show that many types of physical activities can be recognized from field acquired real-world video. Our results also indicate that, with a design of specific motion features in the input vectors, different classifiers can be used successfully with similar performances.
 
Article
In the present paper, we use data mining methods to address two challenges in the sharing and integration of data from electrophysiological (ERP) studies of human brain function. The first challenge, ERP metric matching, is to identify correspondences among distinct summary features ("metrics") in ERP datasets from different research labs. The second challenge, ERP pattern matching, is to align the ERP patterns or "components" in these datasets. We address both challenges within a unified framework. The utility of this framework is illustrated in a series of experiments using ERP datasets that are designed to simulate heterogeneities from three sources: (a) different groups of subjects with distinct simulated patterns of brain activity, (b) different measurement methods, i.e, alternative spatial and temporal metrics, and (c) different patterns, reflecting the use of alternative pattern analysis techniques. Unlike real ERP data, the simulated data are derived from known source patterns, providing a gold standard for evaluation of the proposed matching methods. Using this approach, we demonstrate that the proposed method outperforms well-known existing methods, because it utilizes cluster-based structure and thus achieves finer-grained representation of the multidimensional (spatial and temporal) attributes of ERP data.
 
Article
Receptive fields are commonly used to describe spatial characteristics of sensory neuron responses. They can be extended to characterize temporal or dynamical aspects by mapping neural responses in dynamical state spaces. The state-space receptive field of a neuron is the probability distribution of the dynamical state of the stimulus-generating system conditioned upon the occurrence of a spike. We have computed state-space receptive fields for semicircular canal afferent neurons in the bullfrog (Rana catesbeiana). We recorded spike times during broad-band Gaussian noise rotational velocity stimuli, computed the frequency distribution of head states at spike times, and normalized these to obtain conditional pdfs for the state. These state-space receptive fields quantify what the brain can deduce about the dynamical state of the head when a single spike arrives from the periphery.
 
Article
Chronic pain has profound effects on activity. Previous reports indicate chronic inflammatory conditions result in reduced activity which normalizes upon pain treatment. However, there is little systematic investigation of this process. Rheumatoid arthritis is an autoimmune disorder that causes significant joint pain. The K/BxN serum transfer mouse has been characterized as a model for rheumatoid arthritis and chronic pain. We investigated the activity of mice following K/BxN serum transfer vs. control serum and observed the activity changes following delivery of an NSAID, ketorolac. Previous studies have used running wheels and laser beams to monitor activity; we chose to validate a model using cost-effective infrared sensors on individual cages. Each mouse had its baseline activity obtained, which showed significant variation between individual C57Bl/6 mice. Arthritic mice had significantly decreased activity for only the first 11 nights. Conversely, previous work has shown that these animals display tactile allodynia that persists for at least 45 days. Mice were treated with ketorolac in their drinking water (10mg/kg, 15mg/kg, or 20mg/kg) for nights 6-8. The two highest doses showed significant normalization of activity levels. Four nights after ketorolac was stopped, treated animals were still significantly more active than control. The reversal of the reduced activity provides support that the depression relates to the arthritic pain state of the animal. These results indicate the efficacy of activity monitoring to better investigate behavior in persistent pain states. However, insofar as depressed activity reflects pain and disability, the present work raises questions as to the relevance of the tactile thresholds in defining behaviorally relevant pain states.
 
Article
The tail-withdrawal circuit of Aplysia provides a useful model system for investigating synaptic dynamics. Sensory neurons within the circuit manifest several forms of synaptic plasticity. Here, we developed a model of the circuit and investigated the ways in which depression (DEP) and potentiation (POT) contributed to information processing. DEP limited the amount of motor neuron activity that could be elicited by the monosynaptic pathway alone. POT within the monosynaptic pathway did not compensate for DEP. There was, however, a synergistic interaction between POT and the polysynaptic pathway. This synergism extended the dynamic range of the network, and the interplay between DEP and POT made the circuit responded preferentially to long-duration, low-frequency inputs.
 
Article
Responses of neurons in monkey visual cortex are modulated when attention is directed into the receptive field of the neuron: the gain or sensitivity of the response is increased or the synchronization of the spikes to the local field potential (LFP) is increased. We investigated, using model simulations, whether the synchrony of inhibitory networks could link these observations. We found that, indeed, an increase in inhibitory synchrony could enhance the coherence of the model neurons with the simulated LFP, and could have different effects on the firing rate. When the firing rate vs. current (f-I) response curves saturated at high I, attention yielded a shift in sensitivity; alternatively, when the f-I curves were non-saturating, the most significant effect was on the gain of the response. This suggests that attention may act through changes in the synchrony of inhibitory networks.
 
Article
Aplysia feeding behavior is highly variable from cycle to cycle. In some cycles, when the variability causes a mismatch between the animal's movements and the requirements of the feeding task, the variability makes the behavior unsuccessful. We propose that the behavior is variable nevertheless because the variability serves a higher-order functional purpose. When the animal is faced with a new and only imperfectly known feeding task in each cycle, the variability implements a trial-and-error search through the space of possible feeding movements. Over many cycles, this may be the animal's optimal strategy in an uncertain and changing feeding environment.
 
Article
We have developed a neural system identification method for fitting models to stimulus-response data, where the response is a spike train. The method involves using a general nonlinear optimisation procedure to fit models in the time domain. We have applied the method to model bullfrog semicircular canal afferent neuron responses during naturalistic, broad-band head rotations. These neurons respond in diverse ways, but a simple four parameter class of models elegantly accounts for the various types of responses observed.
 
Article
The pyloric network of crustaceans is a model system for the study of the recovery of function after perturbation/injury of a central pattern-generating network. The network is well characterized anatomically and functionally, yet the cellular mechanism underlying the stabilization or recovery of its activity is not known. In a previous theoretical study long-term activity-dependent regulation of ionic conductances was shown to be sufficient to explain the recovery of rhythmic activity after it is lost due to removal of central input. This model, however, did not capture the complex temporal activity dynamics (bouting) that follows decentralization and that precedes the final stable recovery. Here we build a model of a conditional pacemaker neuron whose ionic conductance levels depend on activity as before, but also includes a slow activity-dependent regulation of Ca2+ uptake (and release). Intracellular Ca2+ sensors, representing enzymatic pathways, regulate the Ca2+ pump activity as well as Ca2+ and K+ conductances. Our model suggests that the activity-dependent regulation of Ca2+ uptake as well as ionic currents interact to generate the complex changes in pyloric activity that follows decentralization. Supported by NIMH 64711 and NSF IBN-0090250.
 
Article
Network plasticity arises in large part due to the effects of exogenous neuromodulators. We investigate the neuromodulatory effects on short-term synaptic dynamics. The synapse from the lateral pyloric (LP) to the pyloric dilator (PD) neuron in the pyloric network of the crab C. borealis has both spike-mediated and non-spike-mediated (graded) components. Previous studies have shown that the graded component of this synapse exhibits short-term depression. Recent results from our lab indicate that in the presence of neuromodulatory peptide proctolin, low-amplitude presynaptic stimuli switch the short-term dynamics of this graded component from depression to facilitation. In this study, we show that this facilitation is correlated with the activation of a presynaptic inward current that is blocked by Mn(2+) suggesting that it is a slowly-accumulating Ca(2+) current. We modify a mechanistic model of synaptic release by assuming that the low-voltage-activating Ca(2+) current in our system is composed of two currents with fast (I(CaF)) and slow (I(CaS)) kinetics. We show that if proctolin adjusts the activation rate of I(CaS), this leads to accumulation of local intracellular Ca(2+) in response to multiple presynaptic voltage stimuli which, in turn, results in synaptic facilitation. Additionally, we assume that proctolin increases the maximal conductances of Ca(2+) currents in the model, consistent with the increased synaptic release found in the experiments. We find that these two presynaptic actions of proctolin in the model are sufficient to describe its actions on the short-term dynamics of the LP to PD synapse.
 
Article
We derive a mathematical theory to explain the subthreshold resonance response of a neuron to synaptic input. The theory shows how a neuron combines information from its intrinsic resonant properties with those of the synapse to determine the neuron's generalized resonance response. Our results show that the maximal response of a postsynaptic neuron can lie between the preferred intrinsic frequency of the neuron and the synaptic resonance frequency. We compare our theoretical results to parallel findings on experiments of the crab pyloric central pattern generator.
 
Article
Independent component analysis (ICA) of functional magnetic resonance imaging (fMRI) data is commonly carried out under the assumption that each source may be represented as a spatially fixed pattern of activation, which leads to the instantaneous mixing model. To allow modeling patterns of spatio-temporal dynamics, in particular, the flow of oxygenated blood, we have developed a convolutive ICA approach: spatial complex ICA applied to frequency-domain fMRI data. In several frequency-bands, we identify components pertaining to activity in primary visual cortex (V1) and blood supply vessels. One such component, obtained in the 0.10 Hz band, is analyzed in detail and found to likely reflect flow of oxygenated blood in V1.
 
Article
It has been unclear whether optimal experimental design accounts of data selection may offer insight into evidence acquisition tasks in which the learner's beliefs change greatly during the course of learning. Data from Rehder and Hoffman's eye movement version of Shepard, Horland and Jenkins' classic concept learning task provide an opportunity to address these issues. We introduce a principled probabilistic concept-learning model that describes the development of subjects' beliefs on that task. We use that learning model, together with a sampling function inspired by theory of optimal experimental design, to predict subjects' eye movements on the active learning version of that task. Results show that the same rational sampling function can predict eye movements early in learning, when uncertainty is high, as well as late in learning when the learner is certain of the true category.
 
Article
Resonance tuning in a model of rhythmic movement is compared when the central pattern generator (CPG) consists of two endogenously bursting or two tonically spiking neurons that are connected with reciprocally inhibitory synapses. The CPG receives inhibitory and/or excitatory position feedback from a linear, one-degree-of-freedom mechanical subsystem. As with previously published results [5, 15], resonance tuning is limited to frequencies that are greater than the intrinsic CPG frequency with endogenously bursting neurons. In contrast, with tonically spiking neurons, the resonance tuning range is expanded to frequencies that are below the intrinsic CPG frequency.
 
Article
Neural models of contextual integration typically incorporate a mean firing rate representation. We examine representation of the full spike count distribution, and its usefulness in explaining contextual integration of color stimuli in primary visual cortex. Specifically, we demonstrate that a factorizable model conditioned on the number of spikes can account for both the onset and sustained portions of the response. We also consider a simplified factorizable model that parametrizes the mean of a Gaussian distribution and incorporates a logistic nonlinearity. The model can account for the sustained response but does not fair as well in accounting for onset nonlinearities. We discuss implications for neural coding.
 
Article
When modulators of neuromuscular function alter the motor neuron spike patterns that elicit muscle contractions, it is predicted that they will also retune correspondingly the connecting processes of the neuromuscular transform. Here we confirm this prediction by analyzing data from the cardiac neuromuscular system of the blue crab. We apply a method that decodes the contraction response to the spike pattern in terms of three elementary building-block functions that completely characterize the neuromuscular transform. This method allows us to dissociate modulator-induced changes in the neuromuscular transform from changes in the spike pattern in the normally operating, essentially unperturbed neuromuscular system.
 
Article
A conductance-based model for synaptic transmission and postsynaptic integration reveals how postsynaptic responses and their variability depend on the number of synaptic inputs. With increasing number of balanced stochastic excitatory and inhibitory inputs, the postsynaptic responses and their variance first increase and then decrease again. This non-linearity can be attributed to an anti-correlation between the total excitatory and inhibitory currents. The anti-correlation, which occurs even though the conductances of the individual synapses vary independently of each other, is determined by the total synaptic conductance and grows with the number of inputs. As the number of inputs increases, the membrane potential comes increasingly closer to the resting level.
 
Article
We implemented an experimentally observed orthogonal arrangement of theta and gamma generation circuitry in septotemporal and lamellar dimensions is a two-dimensional model of hippocampus. The model includes three types of cells: pyramidal, basket, and oriens lacunosum-moleculare (OLM) neurons. In this reduced model, application of continuous electric fields allowed us to switch between theta, gamma and mixed theta-gamma regimes without additional pharmacological manipulation. Electric field effects on individual neurons were modeled based on experimental data. Network simulation results predict a flexible experimental technique, which would employ adaptive subthreshold electric fields to continuously modulate neuronal ensemble activity, and can be used for testing cognitive correlates of oscillatory rhythms as well as for suppressing epileptiform activity.
 
Vertical VOR and OKR neuronal circuit (A) and model (B). In A, FPN and FTN are the floccular (FL) projecting neurons and the FL target neurons in superior vestibular nuclei (SVN), respectively, Y is dorsal Y group, Int. is the inter-neurons in SVN projecting to Y group. P is a FL Purkinje cell and g is a granular cell. PVP is position vestibular pause neurons in medial vestibular nuclei (MVN). MT, MST and DLPN are the middle temporal visual area, the medial superior temporal area and the dorsolateral pontine nuclei, respectively. MN, extraocular motor neurons. LTN, lateral terminal nucleus of the accessory optic system. In B, G ecopy preFL&FL (s), G vestib preFL&FL (s) and G visual preFL&FL (s) are pre-FL/FL subsystems each of which represents a transfer function of preFL/FL efference copy pathway, vestibular pathway and visual pathway, respectively. The three components are added in the FL and form the Purkinje cell SS output. GpostFL(s) represents a transfer function of post-FL pathway which transfers the Purkinje cell simple spike activity to a part of the motor command. G visual nonFL (s) and G vestib nonFL (s) represent transfer functions of non-FL visual and vestibular pathways, respectively. Corresponding neuronal circuit to G visual nonFL (s) is not shown in A. h(t), d(t), r(t), f(t), ecopy(t), and x(t) are head position, optokinetic stimulus position, retinal slip position, FL Purkinje cell simple spike firing pattern, efference copy signal, and eye position, respectively. 
Results of the prediction during VORd, OKR, VORs, VORe at 0.1, 0.5, and 2.5Hz in the same format as in Figure 2 except that all the experimental data are superimposed instead of showing their average and that the stimulus traces are omitted. 
Simulation of the flocculectomy. Eye velocities before (solid lines) and after the flocculectomy (dashed lines) are superimposed. 
Article
The vestibuloocular reflex (VOR) in concert with the optokinetic response (OKR) stabilizes vision during head motion. The VOR system characteristics are both compensatory and adaptively self-calibrated. A model was constructed to aid in the understanding of the roles of the cerebellum and other neuronal sites in the performance and adaptation of the vertical VOR. The model structure was based upon the known neuroanatomy, and model parameters were estimated using experimental data. The model can reproduce and predict eye movements and cerebellar Purkinje cell firing patterns during VOR, OKR, and various visual-vestibular mismatch paradigms.
 
Article
The otolith organs in the vestibular system are excellent detectors of linear accelerations. However, any measurement of linear acceleration is ambiguous between a tilt in a gravitational field and an inertial acceleration. Angelaki et al. have put forward a general hypothesis about how inertial accelerations can be computed based on vestibular signals (J. Neurosci. 19 (1999) 316). We have constructed a realistic, detailed model of the relevant systems to test this hypothesis. The model produces useful predictions about what kinds of neurons should be found in the vestibular nucleus if such a computation is actually performed in the vestibular system. The model is constructed using general principles of neurobiological simulation (J. Neurophys. 84 (2000) 2113).
 
Article
Introducing theta-modulated input into a minimal model of the CA3 region of the hippocampus has significant effects on gamma oscillations. In the absence of theta-modulated input, the gamma oscillations are robust across a range of parameters. Introducing theta-modulated input weakens the gamma oscillations to a power more consistent with power spectra acquired from laboratory animals. With these changes, simulations of the hippocampal model are able to reproduce hippocampal power spectra measured in awake mice.
 
Article
A need is identified to build models of the central nervous system that are semi-complete, applied within multiple contexts to multiple tasks, using methodologies that span multiple levels of abstraction. The issues and constraints in building such models are discussed with respect to completeness, validation, cost, scalability and robustness. An approach currently being explored is described that is suited to the creation of large heterogenous models by small independently collaborating research groups. It is based on a network model interface, a software wrapper that abstracts the interaction between a generic component and a generic framework.
 
Article
We introduce a model for the computation of structure-from-motion based on the physiology of visual cortical areas MT and MST. The model assumes that the perception of depth from motion is related to the firing of a subset of MT neurons tuned to both velocity and disparity. The model's MT neurons are connected to each other laterally to form modulatory receptive-field surrounds that are gated by feedback connections from area MST. This allows the building up of a depth map from motion in area MT, even in absence of disparity in the input. Depth maps from motion and from stereo are combined by a weighted average at a final stage. The model's predictions for the interaction between motion and stereo cues agree with previous psychophysical data, both when the cues are consistent with each other or when they are contradictory. In particular, the model shows nonlinearities as a result of early interactions between motion and stereo before their depth maps are averaged. The two cues interact in a way that represents an alternative to the "modified weak fusion" model of depth-cue combination.
 
Article
We introduce a new correlation-based measure of spike timing reliability. Unlike other measures, it does not require the definition of a posteriori "events". It relies on only one parameter, which relates to the timescale of spike timing precision. We test the measure on surrogate data sets with varying amounts of spike time jitter, and missing or additional spikes, and compare it with a widely used histogram-based measure. The measure is efficient and faithful in characterizing spike timing reliability and produces smaller errors in the reliability estimate than the histogram-based measure based on the same number of trials.
 
Article
Variability of the neuronal spike pattern is usually thought of in terms of the information that the different interspike intervals might be encoding. However, the very presence of the variability can have other kinds of functional significance. Here we consider the example of the B15/B16-ARC neuromuscular system of Aplysia, a model system for the study of neuromuscular modulation and control. We show that variability of motor neuron spike timing at the input to the system penetrates throughout the system, affecting all downstream variables including modulator release, modulator concentrations, modulatory actions, and the contraction of the muscle. Furthermore, not only does the variability penetrate through the system, but it is actually instrumental in maintaining its modulation and contractions at a robust, physiological level.
 
Article
A basic understanding of the relationship between activity of individual neurons and macroscopic electrical activity of local field potentials, or electroencephalogram {(EEG)}, may provide guidance for experimental design in neuroscience, improve development of therapeutic approaches in neurology, and offer opportunities for computer-aided design of brain-computer interfaces. We study the relationship between resonant properties of neurons and network oscillations in a computational model of neocortex. Our findings suggest that resonance is associated with subthreshold oscillation of neurons. This subthreshold behavior affects spike timing and plays a significant role in the generation of the network's extracellular currents reflected in the {EEG.
 
Conference Paper
One central issue in a long-tail online marketplace such as eBay is to automatically put user self-input items into a catalog in real time. This task is extremely challenging when the inventory scales up, the items become ephemeral, and the user input remains noisy. Indeed, catalog learning has emerged as a key technical property for other major online ecommerce applications including search and recommendation. We formulate the item cataloging task as a Bayesian classification problem, which shall scale well in very large data set and have good online prediction performance. The inherent data sparseness issue, especially for those tail categories, is key to the overall model performance. We address the data sparseness issue by adapting statistically sound smoothing methods well studied in language modeling tasks. However, there are data characteristics specific to the ecommerce domain, including short yet focused item description, very large and hierarchical catalog taxonomy, and highly skewed distribution over types of items. We investigate these domain-specific regularities empirically, and report practically significant results with real-world true-scale data.
 
Conference Paper
A new approach to the design of S-model ergodic reinforcement learning algorithms is introduced. The new scheme utilizes a stochastic estimator and can operate in nonstationary environments with high accuracy and high adaptation rate. The performance of the presented stochastic estimator learning automation (SELA) is superior over previous well-known S-model ergodic schemes. Furthermore it is proved that SELA is ε-optimal in every S-model random environment
 
Conference Paper
Appearances of objects lie in high-dimensional spaces. For a given recognition task, feature selection aims to select most effective features in order to reduce the recognition cost and improve recognition accuracy. Feature selection can be achieved by a bottom-up scheme, e.g., using spatial information, or a top-down scheme, e.g., using class information. In this paper, we propose a model to integrate spatial and discriminant influence for appearance based recognition, where locality oriented Fisher score is introduced to estimate the discriminant influence. We use Lipschitz regularity to construct image representation. We present a case study of embryo stage recognition to test the performance of the proposed method. We also obtain new insights on the comparison between spatial and discriminant influence.
 
Conference Paper
In this paper, we present a new method to recognize object class based on local appearance features and context information. At first, local descriptors of object class appearance are clustered, then part classifiers are trained to select the most distinctive image patches and visual context information around them are extracted to keep the robustness to object occlusion and background clutter. Finally general probabilistic models are built to implement image classification by integrating the context information with local scale-invariant appearance characteristics. Compared with previous work, we obtain a better classification with limited and unnormalized training samples. Experiment results show that the proposed method can outperform other previous methods even under large scale object classes, therefore the significance of appearance-based discriminative part classifiers is demonstrated and confirmed.
 
Conference Paper
A new multidirectional associative memory (MAM), named combined multi-winner MAM (CMW-MAM) using distributed representation paradigm, is proposed and the association properties for storage of complex memory structure are studied. The conventional MAM stores and recalls information directly, however, the proposed CMW-MAM stores and recalls information hierarchically by the distributed representation which are formed automatically in the networks. The proposed CMW-MAM can handle analog information and can store and recall more complex memory structures. The robustness against incomplete information and internal damage is very strong
 
Conference Paper
We review a recently developed engineering control approach to attention, presenting detailed attention control function assignments to the wealth of brain modules experimentally observed. The control system is extended to include biasing by emotional valence, with qualitative analysis given of a range of emotion paradigms and more detailed simulation described for two further paradigms. The implications of these results for better understanding of the interaction of emotion and attention concludes the paper, and in, particular gives a possible resolution of the question as to unaware versus aware processing of emotional material.
 
Conference Paper
In this paper, the multigrid-based fuzzy system (MGFS) approach is applied for the CATS time series prediction benchmark. The MGFS architecture overcomes the problem inherent to all grid-based fuzzy systems when dealing with high dimensional input data, thus keeping low computational cost and high performance. A greedy algorithm for MGFS structure identification allows to perform the input variable selection for the time series prediction problem, while identifying the pseudo-optimal architecture according to the provided dataset.
 
Conference Paper
The current generation of nonmodular neural network classifiers is unable to cope with classification problems which have a wide range of overlap among classes. This is due to the high coupling among the networks' hidden nodes. We propose the Cooperative Modular Neural Network (CMNN) architecture, which deals with different levels of overlap in different modules. The modules share their information and cooperate in taking a global classification decision through voting. Moreover, special modules are dedicated to resolve high overlaps in the input-space. The performance of the new model outperforms that of the nonmodular alternative when when applied to ten famous benchmark classification problems
 
Conference Paper
We propose a topological local principal component analysis (PCA) in help of Kohonen's self-organizing maps (SOM). The topological local PCA describes one cluster by one neuron such that it is capable of exploiting both the global topological structure and each local cluster structure. We also investigate a newly proposed self-organizing strategy that can enhance the learning speed, as well as an alternative Stiefel manifold based algorithm to ensure the orthonormality constraint of the local PCA.
 
Conference Paper
We address the problem of blind separation of convolutive mixtures of spatially and temporally independent sources modeled with mixtures of Gaussians. We present an EM algorithm to compute maximum likelihood estimates of both the separating filters and the source density parameters, whereas, in the state-of-the-art, separating filters are usually estimated with gradient descent techniques. The use of the EM algorithm, as opposed to the usual gradient descent techniques, does not require the empirical tuning of a learning rate, and thus can be expected to provide a more stable convergence. Besides, we show how multichannel autoregressive spectral estimation techniques can be used in order to initialize the EM algorithm properly. We demonstrate the efficiency of our EM algorithm together with the proposed initialization scheme by reporting on simulations with artificial mixtures.
 
Conference Paper
Compressive Sensing Deconvolution (CS Deconvolution) is a new challenge problem encountered in a wide variety of image processing fields. A compound variational regularization model which combined total variation and curve let-based sparsity prior is proposed to recovery blurred image from compressive measurements. We propose a novel fast algorithm using variable-splitting and Dual Douglas-Rachford operator splitting methods. Experiments demonstrate our proposed algorithm can obtain high-resolution data from highly incomplete measurements.
 
Conference Paper
The aim of semantic segmentation is to assign each pixel a semantic label. Numerous methods for semantic segmentation have been proposed in recent years and most of them chose pixel or super pixel as the processing primitives. However, as the information contained in a pixel or a super pixel is not discriminative enough, the outputs of these algorithms are usually not object consistent. To tackle this problem, we introduce the concept of object-like regions as a new and higher level processing primitive. We first experimentally showed that using object-like regions as processing primitives can boost semantic segmentation accuracy, and then proposed a novel method to produce object-like regions by integrating state-of art low-level segmentation algorithms with typical semantic segmentation algorithms through a novel semantic feature feedback mechanism. We present experimental results on the publicly available image understanding database MSRC21 and show that the new method can achieve state of the art semantic segmentation results with far fewer processing primitives.
 
Conference Paper
A piecewise linear neural network (PLNN) is discussed which maps N-dimensional input vectors into M-dimensional output vectors. A convergent algorithm for designing the PLNN from training data is described The design algorithm is based on a variation of backtracking algorithm known as the `branch-and-bound' method. The performance of the PLNN is compared with that of a multilayer perceptron (MLP) of equivalent size. The results show that the PLNN is capable of performing as well as an equivalent MLP
 
Conference Paper
An approach to reducing the computation time taken by fast neural nets for the searching process is presented. The principle of the divide and conquer strategy is applied through image decomposition. Each image is divided into small in size sub-images and then each one is tested separately using a fast neural network Compared to conventional and fast neural networks, experimental results show that a speed up ratio is achieved when applying this technique to locate human faces automatically in cluttered scenes. Furthermore, faster face detection is obtained by using parallel processing techniques to test the resulting sub-images at the same time using the same number of fast neural networks. Moreover, the problem of sub-image centering and normalization in the Fourier space is solved
 
Conference Paper
The objectives of this paper are to derive a momentum term in the Kohonen's self-organizing feature map algorithm theoretically and to show the effectiveness of the term by computer simulations. We will derive a self-organizing feature map algorithm having the momentum term through the following assumptions: 1) The cost function is E<sup>n</sup>=Σ<sub>μ</sub><sup>n</sup>α<sup>n-μ</sup>E<sub>μ</sub>, where E<sub>μ</sub> is the modified Lyapunov function originally proposed by Ritter and Schulten (1988, 1992) at the μth learning time and α is the momentum coefficient. 2) The latest weights are assumed in calculating the cost function E<sup>n</sup>. According to our simulations, it has shown that the momentum term in the self-organizing feature map can considerably contribute to the acceleration of the convergence.
 
Article
Describing neuroanatomical circuitry at the systems level requires the use of databases to collate published descriptions of neuroanatomical data (Burns, D.Phil. Thesis, Physiology Department, Oxford University, 1997; Felleman and Van Essen, Cerebral Cortex, 1, 1991, 1–47; Scanell et al., J. Neurosci. 15, 1995, 1463–83; Young, Proc. R. Soc. London Ser B, 252, 1993, 13–8). These data are then analyzed with computational techniques. These methods do not address the problem that qualitative neuroanatomical descriptions can be interpreted in different ways, and rely on the collator's skill to produce the correct interpretation. I describe a knowledge-base management system called “NeuroScholar”, designed to store multiple interpretations of neuroanatomical tract-tracing data in a neuroanatomically consistent framework. I illustrate how this system may be used in conjunction with data-mining analyses.
 
Article
This article introduces the “CAM-Brain Machine” (CBM), an FPGA-based piece of hardware that implements a genetic algorithm (GA) to evolve a cellular automata (CA)-based neural network circuit module (of approximately 1000 neurons) in seconds (i.e., a complete run of a GA, with 10,000s of circuit growths and performance evaluations). Up to 65,000 of these modules (each of which is evolved with a specified function) can be downloaded into a gigabyte of RAM space, and interconnected according to specified artificial brain architectures. This RAM, containing an “artificial brain” with up to 75 million neurons, is then updated by the CBM at a rate of 130 billion CA cells per second. Such speeds should enable real-time control of robots and hopefully the birth of a new research field that we call “brain building”. The first such artificial brain (to be built by STARLAB starting in 2001) will be used to control the behaviors of a life-sized robot kitten called “Robokitty”.
 
Article
A method for detecting relationships between single-unit spike trains and local field potentials (LFPs) was developed and applied to recordings from rat cerebellum. LFPs were repeatedly filtered with a shifting frequency window. The resulting traces were transformed into peak-time point processes for comparison with spike trains using “relative-phase” analysis (Biol. Cybernet., in press). Discharge of some Purkinje cells was phase-related to LFP oscillations in the 2– and 30– frequency ranges. This analysis method revealed hidden coherency between spike-trains and LFPs. The findings suggest that cerebellar activity is, to some extent, temporally organized according to both slow and fast rhythms.
 
Article
Autoregressive integrated moving average (ARIMA) is one of the popular linear models in time series forecasting during the past three decades. Recent research activities in forecasting with artificial neural networks (ANNs) suggest that ANNs can be a promising alternative to the traditional linear methods. ARIMA models and ANNs are often compared with mixed conclusions in terms of the superiority in forecasting performance. In this paper, a hybrid methodology that combines both ARIMA and ANN models is proposed to take advantage of the unique strength of ARIMA and ANN models in linear and nonlinear modeling. Experimental results with real data sets indicate that the combined model can be an effective way to improve forecasting accuracy achieved by either of the models used separately.
 
Article
Non-invasive techniques such as magnetic resonance spectroscopy (MRS) are often required for assisting the diagnosis of tumours. Radiologists are not always accustomed to make sense of the biochemical information provided by MRS and they may benefit from computer-based support in their decision making. The high dimensionality of the MR spectra obscures atypical aspects of the data that may jeopardize their classification. In this study, we describe a method to overcome this problem that combines nonlinear dimensionality reduction, outlier detection, and expert opinion. MR spectra subsequently undergo a feature selection process followed by classification. The impact of outlier removal on classification performance is assessed.
 
Article
In this study, we concentrate on the fundamentals and essential development issues of logic-driven constructs of fuzzy neural networks. These networks, referred to as logic-oriented neural networks, constitute an interesting conceptual and computational framework that greatly benefits from the establishment of highly synergistic links between the technology of fuzzy sets (or granular computing, being more general) and neural networks.The most essential advantages of the proposed networks are twofold. First, the transparency of neural architectures becomes highly relevant when dealing with the mechanisms of efficient learning. Here the learning is augmented by the fact that domain knowledge could be easily incorporated in advance prior to any learning. This becomes possible given the compatibility between the architecture of the problem and the induced topology of the neural network. Second, once the training has been completed, the network can be easily interpreted and thus it directly translates into a series of truth-quantifiable logic expressions formed over a collection of information granules.The design process of the logic networks synergistically exploits the principles of information granulation, logic computing and underlying optimization including those biologically inspired techniques (such as particle swarm optimization, genetic algorithms and alike). We elaborate on the existing development trends, present key methodological pursuits and algorithms. In particular, we show how the logic blueprint of the networks is supported by the use of various constructs of fuzzy sets including logic operators, logic neurons, referential operators and fuzzy relational constructs.
 
Article
Variations in electrical load are, among other things, hour of the day dependent, introducing a dilemma for the forecaster: whether to partition the data and use a separate model for each hour of the day (the parallel approach), or use a single model (the sequential approach). This paper examines which approach is appropriate for forecasting hourly electrical load in Ireland. It is found that, with the exception of some hours of the day, the sequential approach is superior. The final solution however, uses a combination of linear sequential and parallel neural models in a multi-time scale formulation.
 
Top-cited authors
Guang-Bin Huang
  • Nanyang Technological University
Chee Kheong Siew
  • Nanyang Technological University
Yurong Liu
  • Yangzhou University
Zidong Wang
  • Harbin University of Science and Technology
Nianyin Zeng
  • Xiamen University