Self-organizing neural projections.

The Self-Organizing Map (SOM) algorithm was developed for the creation of abstract-feature maps. It has been accepted widely as a data-mining tool, and the principle underlying it may also explain how the feature maps of the brain are formed. However, it is not correct to use this algorithm for a model of pointwise neural projections such as the somatotopic maps or the maps of the visual field, first of all, because the SOM does not transfer signal patterns: the winner-take-all function at its output only defines a singular response. Neither can the original SOM produce superimposed responses to superimposed stimulus patterns. This presentation introduces a new self-organizing system model related to the SOM that has a linear transfer function for patterns and combinations of patterns all the time. Starting from a randomly interconnected pair of neural layers, and using random mixtures of patterns for training, it creates a pointwise-ordered projection from the input layer to the output layer. If the input layer consists of feature detectors, the output layer forms a feature map of the inputs.     

Integrating support vector machines and neural networks.

Support vector machines (SVMs) are a powerful technique developed in the last decade to effectively tackle classification and regression problems. In this paper we describe how support vector machines and artificial neural networks can be integrated in order to classify objects correctly. This technique has been successfully applied to the problem of determining the quality of tiles. Using an optical reader system, some features are automatically extracted, then a subset of the features is determined and the tiles are classified based on this subset.     

Self-organizing continuous attractor networks and motor function.

Motor skill learning may involve training a neural system to automatically perform sequences of movements, with the training signals provided by a different system, used mainly during training to perform the movements, that operates under visual sensory guidance. We use a dynamical systems perspective to show how complex motor sequences could be learned by the automatic system. The network uses a continuous attractor network architecture to perform path integration on an efference copy of the motor signal to keep track of the current state, and selection of which motor cells to activate by a movement selector input where the selection depends on the current state being represented in the continuous attractor network. After training, the correct motor sequence may be selected automatically by a single movement selection signal. A feature of the model presented is the use of 'trace' learning rules which incorporate a form of temporal average of recent cell activity. This form of temporal learning underlies the ability of the networks to learn temporal sequences of behaviour. We show that the continuous attractor network models developed here are able to demonstrate the key features of motor function. That is, (i) the movement can occur at arbitrary speeds; (ii) the movement can occur with arbitrary force; (iii) the agent spends the same relative proportions of its time in each part of the motor sequence; (iv) the agent applies the same relative force in each part of the motor sequence; and (v) the actions always occur in the same sequence.     

Identification of motifs with insertions and deletions in protein sequences using self-organizing neural networks.

The problem of motif identification in protein sequences has been studied for many years in the literature. Current popular algorithms of motif identification in protein sequences face two difficulties, high computational cost and the possibility of insertions and deletions. In this paper, we provide a new strategy that solve the problem more efficiently. We develop a self-organizing neural network structure with multiple levels of subnetworks to make an intelligent classification of the subsequences obtained from protein sequences. We maintain a low computational complexity through the use of this multi-level structure so that the classification of each subsequence is performed with respect to a small subspace of the whole input space. The new definition of pairwise distance between motif patterns provided in this paper can deal with up to two insertions/deletions allowed in a motif, while other existing algorithm can only deal with one insertion or deletion. We also maintain a high reliability using our self-organizing neural network since it will grow as needed to make sure all input patterns are considered and are given the same amount of attention. Simulation results show that our algorithm significantly outperforms existing algorithms in both accuracy and reliability aspects.     

Two distinct neural networks support the mapping of meaning to a novel word

Children can learn the meaning of a new word from context during normal reading or listening, without any explicit instruction. It is unclear how such meaning acquisition is supported and achieved in human brain. In this functional magnetic resonance imaging (fMRI) study we investigated neural networks supporting word learning with a functional connectivity approach. Participants were exposed to a new word presented in two successive sentences and needed to derive the meaning of the new word. We observed two neural networks involved in mapping the meaning to the new word. One network connected the left inferior frontal gyrus (LIFG) with the middle frontal gyrus (MFG), medial superior frontal gyrus, caudate nucleus, thalamus, and inferior parietal lobule. The other network connected the left middle temporal gyrus (LMTG) with the MFG, anterior and posterior cingulate cortex. The LIFG network showed stronger interregional interactions for new than real words, whereas the LMTG network showed similar connectivity patterns for new and real words. We proposed that these two networks support different functions during word learning. The LIFG network appears to select the most appropriate meaning from competing candidates and to map the selected meaning onto the new word. The LMTG network may be recruited to integrate the word into sentential context, regardless of whether the word is real or new. The LIFG and the LMTG networks share a common node, the MFG, suggesting that these two networks communicate in working memory. Hum Brain Mapp, 2011. © 2010 Wiley‐Liss, Inc.     

Distinct neural networks support the mere ownership effect under different motivational contexts

The “mere ownership effect” refers to individuals’ tendency to evaluate objects they own more favorably than comparable objects they do not own. There are numerous behavioral demonstrations of the mere ownership effect, but the neural mechanisms underlying the expression of this self-positivity bias during the evaluation of self-associated objects have not been identified. The present study aimed to identify the neurobiological expression of the mere ownership effect and to assess the potential influence of motivational context. During fMRI scanning, participants made evaluations of objects after ownership had been assigned under the presence or absence of self-esteem threat. In the absence of threat, the mere ownership effect was associated with brain regions implicated in processing personal/affective significance and valence (ventromedial prefrontal cortex [vMPFC], ventral anterior cingulate cortex [vACC], and medial orbitofrontal cortex [mOFC]). In contrast, in the presence of threat, the mere ownership effect was associated with brain regions implicated in selective/inhibitory cognitive control processes (inferior frontal gyrus [IFG], middle frontal gyrus [MFG], and lateral orbitofrontal cortex [lOFC]). These findings indicate that depending on motivational context, different neural mechanisms (and thus likely different psychological processes) support the behavioral expression of self-positivity bias directed toward objects that are associated with the self.     

An approach to guaranteeing generalisation in neural networks.

A novel approach to generalisation is presented that is able, under certain circumstances, to guarantee the generalisation to binary-output data for which no targets have been given. The basis of the guarantee is the recognition of a persistent global minimum error solution. An empirical test for whether the guarantee holds is provided which uses a technique called target reversal. The technique employs two neural networks whose convergence using opposing targets signals validity of the guarantee.     

Information complexity of neural networks.

This paper studies the question of lower bounds on the number of neurons and examples necessary to program a given task into feed forward neural networks. We introduce the notion of information complexity of a network to complement that of neural complexity. Neural complexity deals with lower bounds for neural resources (numbers of neurons) needed by a network to perform a given task within a given tolerance. Information complexity measures lower bounds for the information (i.e. number of examples) needed about the desired input-output function. We study the interaction of the two complexities, and so lower bounds for the complexity of building and then programming feed-forward nets for given tasks. We show something unexpected a priori--the interaction of the two can be simply bounded, so that they can be studied essentially independently. We construct radial basis function (RBF) algorithms of order n3 that are information-optimal, and give example applications.     

Self-organizing neural network analyses of cardiac data in depression

OBJECTIVE: To determine if an unsupervised self-organizing neural network could create a clinically meaningful distinction of 'depression' versus 'no depression' based on cardiac time-series data. DESIGN: A self-organizing map (SOM) was used to separate the time-series of 84 subjects into groups based on characteristics of the data alone. MATERIALS AND METHODS: Analyses included natural log transformations and two types of filtering to enhance characteristics of the data as well as classifications of unprocessed data. A Pearson chi(2) analysis was performed to determine if the SOM groups bore any relation to the binary clinical groups. RESULTS: Overall correct SOM classifications ranged from 54 to 70.2% with two classifications being clinically meaningful. CONCLUSIONS: SOM classifications of cardiac time-series data with enhanced ultradian variations and cardiac data recorded around the interval when a person was in bed were useful in differentiating clinically meaningful subgroups with and without depression.     

The use of artificial neural networks in decision support in cancer: a systematic review.

Artificial neural networks have featured in a wide range of medical journals, often with promising results. This paper reports on a systematic review that was conducted to assess the benefit of artificial neural networks (ANNs) as decision making tools in the field of cancer. The number of clinical trials (CTs) and randomised controlled trials (RCTs) involving the use of ANNs in diagnosis and prognosis increased from 1 to 38 in the last decade. However, out of 396 studies involving the use of ANNs in cancer, only 27 were either CTs or RCTs. Out of these trials, 21 showed an increase in benefit to healthcare provision and 6 did not. None of these studies however showed a decrease in benefit. This paper reviews the clinical fields where neural network methods figure most prominently, the main algorithms featured, methodologies for model selection and the need for rigorous evaluation of results.     

Supramodal neural networks support top‐down processing of social signals

The perception of facial and vocal stimuli is driven by sensory input and cognitive top‐down influences. Important top‐down influences are attentional focus and supramodal social memory representations. The present study investigated the neural networks underlying these top‐down processes and their role in social stimulus classification. In a neuroimaging study with 45 healthy participants, we employed a social adaptation of the Implicit Association Test. Attentional focus was modified via the classification task, which compared two domains of social perception (emotion and gender), using the exactly same stimulus set. Supramodal memory representations were addressed via congruency of the target categories for the classification of auditory and visual social stimuli (voices and faces). Functional magnetic resonance imaging identified attention‐specific and supramodal networks. Emotion classification networks included bilateral anterior insula, pre‐supplementary motor area, and right inferior frontal gyrus. They were pure attention‐driven and independent from stimulus modality or congruency of the target concepts. No neural contribution of supramodal memory representations could be revealed for emotion classification. In contrast, gender classification relied on supramodal memory representations in rostral anterior cingulate and ventromedial prefrontal cortices. In summary, different domains of social perception involve different top‐down processes which take place in clearly distinguishable neural networks.     

Training neural networks to be insensitive to weight random variations.

Neural network weights are subject to errors caused by technological tolerances when implemented in digital or analog hardware. Since these random variations are unavoidable and unpredictable, they can seriously affect the expected performances. This work proposes a learning algorithm that takes weight tolerances into account and guarantees a low sensitivity to them. Some experimental results show the validity of the suggested approach.     

Evolving neural networks to identify bent-double galaxies in the FIRST survey.

The FIRST (Faint Images of the Radio Sky at Twenty-cm) survey is an ambitious project scheduled to cover 10,000 square degrees of the northern and southern galactic caps. Until recently, astronomers associated with FIRST identified radio-emitting galaxies with a bent-double morphology through a visual inspection of images. Besides being subjective, prone to error and tedious, this manual approach is becoming increasingly infeasible: upon completion, FIRST will include almost a million galaxies. This paper describes the application of six methods of evolving neural networks (NNs) with genetic algorithms (GAs) to the identification of bent-double galaxies. The objective is to demonstrate that GAs can successfully address some common problems in the application of NNs to classification problems, such as training the networks, choosing appropriate network topologies, and selecting relevant features. We measured the overall accuracy of the networks using the arithmetic and geometric means of the accuracies on bent and non-bent galaxies. Most of the combinations of GAs and NNs perform equally well on our data, but using GAs to select feature subsets produces the best results, reaching accuracies of 90% using the arithmetic mean and 87% with the geometric mean. The networks found by the GAs were more accurate than hand-designed networks and decision trees.     

On impulsive autoassociative neural networks.

Many systems existing in physics, chemistry, biology, engineering, and information science can be characterized by impulsive dynamics caused by abrupt jumps at certain instants during the process. These complex dynamical behaviors can be modeled by impulsive differential systems or impulsive neural networks. This paper formulates and studies a new model of impulsive autoassociative neural networks. Several fundamental issues, such as global exponential stability and existence and uniqueness of equilibria of such neural networks, are established.     

Electroencephalogram processing using neural networks.

The electroencephalogram (EEG), a highly complex signal, is one of the most common sources of information used to study brain function and neurological disorders. More than 100 current neural network applications dedicated to EEG processing are presented. Works are categorized according to their objective (sleep analysis, monitoring anesthesia depth, brain-computer interface, EEG artifact detection, EEG source-based localization, etc.). Each application involves a specific approach (long-term analysis or short-term EEG segment analysis, real-time or time delayed processing, single or multiple EEG-channel analysis, etc.), for which neural networks were generally successful. The promising performances observed are demonstrative of the efficiency and efficacy of systems developed. This review can aid researchers, clinicians and implementors to understand up-to-date interest in neural network tools for EEG processing. The extended bibliography provides a database to assist in possible new concepts and idea development.     

Dynamics of periodic delayed neural networks.

This paper formulates and studies a model of periodic delayed neural networks. This model can well describe many practical architectures of delayed neural networks, which is generalization of some additive delayed neural networks such as delayed Hopfield neural networks and delayed cellular neural networks, under a time-varying environment, particularly when the network parameters and input stimuli are varied periodically with time. Without assuming the smoothness, monotonicity and boundedness of the activation functions, the two functional issues on neuronal dynamics of this periodic networks, i.e. the existence and global exponential stability of its periodic solutions, are investigated. Some explicit and conclusive results are established, which are natural extension and generalization of the corresponding results existing in the literature. Furthermore, some examples and simulations are presented to illustrate the practical nature of the new results.     

Exponential stability of Cohen-Grossberg neural networks.

Exponential stabilities of the Cohen-Grossberg neural network with and without delays are analyzed. By Liapunov functions/functionals, sufficient conditions are obtained for general exponential stability, while by using a comparison result from the theory of monotone dynamical systems, componentwise exponential stability is also discussed. All results are established without assuming any symmetry of the connection matrix, and the differentiability and monotonicity of the activation functions.     

Application of neural networks to fuzzy control

This paper gives a possible application of neural networks to fuzzy control. In fuzzy control a set of linguistic rules are given and by specifying a method of fuzzy reasoning and defuzzification an input-output relation is obtained. Fuzzy controllers thus obtained are usually irregular, and are not necessarily what experts expect. It is sometimes difficult to implement such controllers when processing time is limited. Here we attempt to dissolve such drawbacks using neural networks that can learn these input-output maps. We show that good neuro-controller can be obtained for an inverted pendulum system. The structure of the neuro-controller is simple, and hence analysis and implementation are easy. We discuss the stability of the system and confirm our results by experiments.     

A new algorithm to design compact two-hidden-layer artificial neural networks.

This paper describes the cascade neural network design algorithm (CNNDA), a new algorithm for designing compact, two-hidden-layer artificial neural networks (ANNs). This algorithm determines an ANN's architecture with connection weights automatically. The design strategy used in the CNNDA was intended to optimize both the generalization ability and the training time of ANNs. In order to improve the generalization ability, the CNDDA uses a combination of constructive and pruning algorithms and bounded fan-ins of the hidden nodes. A new training approach, by which the input weights of a hidden node are temporarily frozen when its output does not change much after a few successive training cycles, was used in the CNNDA for reducing the computational cost and the training time. The CNNDA was tested on several benchmarks including the cancer, diabetes and character-recognition problems in ANNs. The experimental results show that the CNNDA can produce compact ANNs with good generalization ability and short training time in comparison with other algorithms.     

Comparing error minimized extreme learning machines and support vector sequential feed-forward neural networks

Recently, error minimized extreme learning machines (EM-ELMs) have been proposed as a simple and efficient approach to build single-hidden-layer feed-forward networks (SLFNs) sequentially. They add random hidden nodes one by one (or group by group) and update the output weights incrementally to minimize the sum-of-squares error in the training set. Other very similar methods that also construct SLFNs sequentially had been reported earlier with the main difference that their hidden-layer weights are a subset of the data instead of being random. These approaches are referred to as support vector sequential feed-forward neural networks (SV-SFNNs), and they are a particular case of the sequential approximation with optimal coefficients and interacting frequencies (SAOCIF) method. In this paper, it is firstly shown that EM-ELMs can also be cast as a particular case of SAOCIF. In particular, EM-ELMs can easily be extended to test some number of random candidates at each step and select the best of them, as SAOCIF does. Moreover, it is demonstrated that the cost of the computation of the optimal output-layer weights in the originally proposed EM-ELMs can be improved if it is replaced by the one included in SAOCIF. Secondly, we present the results of an experimental study on 10 benchmark classification and 10 benchmark regression data sets, comparing EM-ELMs and SV-SFNNs, that was carried out under the same conditions for the two models. Although both models have the same (efficient) computational cost, a statistically significant improvement in generalization performance of SV-SFNNs vs. EM-ELMs was found in 12 out of the 20 benchmark problems.