Artificial Neural Networks and Machine Learning – ICANN 2014

In this paper a novel recurrent neural network (RNN) model for gradient-based sequence learning is introduced. The presented dynamic cortex memory (DCM) is an extension of the well-known long short term memory (LSTM) model. The main innovation of the DCM is the enhancement of the inner interplay of the gates and the error carousel due to several new and trainable connections. These connections enable a direct signal transfer from the gates to one another. With this novel enhancement the networks are able to converge faster during training with back-propagation through time (BPTT) than LSTM under the same training conditions. Furthermore, DCMs yield better generalization results than LSTMs. This behaviour is shown for different supervised problem scenarios, including storing precise values, adding and learning a context-sensitive grammar.

In the present study, we demonstrate the learning and recognition capabilities of our recently proposed recurrent neural network (RNN) model called stochastic continuous-time RNN (S-CTRNN). S-CTRNN can learn to predict not only the mean but also the variance of the next state of the learning targets. The network parameters consisting of weights, biases, and initial states of context neurons are optimized through maximum likelihood estimation (MLE) using the gradient descent method. First, we clarify the essential difference between the learning capabilities of conventional CTRNN and S-CTRNN by analyzing the results of a numerical experiment in which multiple fluctuating temporal patterns were used as training data, where the variance of the Gaussian noise varied among the patterns. Furthermore, we also show that the trained S-CTRNN can recognize given fluctuating patterns by inferring the initial states that can reproduce the patterns through the same MLE scheme as that used for network training.

We introduce a regularization technique to improve system identification for dual-task learning with recurrent neural networks. In particular, the method is introduced using the Factored Tensor Recurrent Neural Networks first presented in [1]. Our goal is to identify a dynamical system with few available observations by augmenting them with data from a sufficiently observed similar system. In our previous work, we discovered that the model accuracy degrades whenever little data of the system of interest is available. The presented regularization term in this work allows to significantly reduce the model error thereby improving the exploitation of knowledge of the well observed system. This scenario is crucial in many real world applications, where data efficiency plays an important role. We motivate the problem setting and our regularized dual-task learning approach by industrial use cases, e.g. gas or wind turbine modeling for optimization and monitoring. Then, we formalize the problem and describe our regularization term by which the learning objective of the Factored Tensor Recurrent Neural Network is extended. Finally, we demonstrate its effectiveness on the cart-pole and mountain car benchmarks.

In the present paper we investigate influence of IP tuning of Echo state network (ESN) reservoir on the overall behavior of the on-line trained adaptive critic network. The experiments were done using Adaptive Critic Design (ACD) scheme with on-line trainable ESN critic for real time control of a mobile laboratory robot. Comparison of behavior of ESN critics trained with and without IP tuning showed that IP algorithm improved critic behavior significantly. It was observed that IP tuning prevents uncontrolled increase of reservoir output weights during on-line training.

In previous research a model for thematic role assignment (

θ

RARes) was proposed, using the Reservoir Computing paradigm. This language comprehension model consisted of a recurrent neural network (RNN) with fixed random connections which models distributed processing in the prefrontal cortex, and an output layer which models the striatum. In contrast to this previous batch learning method, in this paper we explored a more biological learning mechanism. A new version of the model (i-

θ

RARes) was developed that permitted incremental learning, at each time step. Learning was based on a stochastic gradient descent method. We report here results showing that this incremental version was successfully able to learn a corpus of complex grammatical constructions, reinforcing the neurocognitive plausibility of the model from a language acquisition perspective.

Reservoir computing provides a promising approach to efficient training of recurrent neural networks, by exploiting the computational properties of the reservoir structure. Various approaches, ranging from suitable initialization to reservoir optimization by training have been proposed. In this paper we take a closer look at short-term memory capacity, introduced by Jaeger in case of echo state networks. Memory capacity has recently been investigated with respect to criticality, the so called edge of chaos, when the network switches from a stable regime to an unstable dynamic regime. We calculate memory capacity of the networks for various input data sets, both random and structured, and show how the data distribution affects the network performance. We also investigate the effect of reservoir sparsity in this context.

The echo-state condition names an upper limit for the hidden layer connectivity in recurrent neural networks. If the network is below this limit there is an injective, continuous mapping from the recent input history to the internal state of the network. Above the network becomes chaotic, the dependence on the initial state of the network may never be washed out. I focus on the biological relevance of echo state networks with a critical connectivity strength at the separation line between these two conditions and discuss some related biological findings, i.e. there is evidence that the neural connectivity in cortical slices is tuned to a critical level. In addition, I propose a model that makes use of a special learning mechanism within the recurrent layer and the input connectivity. Results show that after adaptation indeed traces of single unexpected events stay for a longer time period than exponential in the network.

Understanding the dynamical and computational capabilities of neural models represents an issue of central importance. In this context, recent results show that interactive evolving recurrent neural networks are super-Turing, irrespective of whether their synaptic weights are rational or real. We extend these results by showing that interactive evolving recurrent neural networks are not only super-Turing, but also capable of simulating any other possible interactive deterministic system. In this sense, interactive evolving recurrent neural networks represents a super-Turing universal model of computation, irrespective of whether their synaptic weights are rational or real.

Slow adaption processes, like synaptic and intrinsic plasticity, abound in the brain and shape the landscape for the neural dynamics occurring on substantially faster timescales. At any given time the network is characterized by a set of internal parameters, which are adapting continuously, albeit slowly. This set of parameters defines the number and the location of the respective adiabatic attractors. The slow evolution of network parameters hence induces an evolving attractor landscape, a process which we term attractor metadynamics. We study the nature of the metadynamics of the attractor landscape for several continuous-time autonomous model networks. We find both first- and second-order changes in the location of adiabatic attractors and argue that the study of the continuously evolving attractor landscape constitutes a powerful tool for understanding the overall development of the neural dynamics.

The digital spike-phase map is a simple digital dynamical system that can generate various spike-trains. In order to approach systematic analysis of the steady and transient states, four basic feature quantities are presented. Using the quantities, we analyze an example based on the bifurcating neuron with triangular base signal and consider basic four cases of the spike-train dynamics.

Nonnegative Matrix Factorization (NMF) is more and more frequently used for analyzing large-scale nonnegative data, where the number of samples and/or the number of observed variables is large. In the paper, we discuss two applications of the row-action projections in the context of learning latent factors from large-scale data. First, we show that they can be efficiently used for improving the on-line learning in dynamic NMF. Next, they can also considerably reduce the computational complexity of the optimization algorithms used for factor learning from strongly redundant data. The experiments demonstrate high efficiency of the proposed methods.

In this paper, we extract the core idea of state perturbation from

Hopfield-type

neural networks and define state perturbation formulas to describe the general way of optimization methods. Departing from the core idea and the formulas, we propose a novel optimization method related to neural network structure, named structure perturbation optimization. Our method can produce a structure transforming process to retrain

Hopfield-type

neural networks to get better problem-solving ability. Experiments validate that our method effectively helps

Hopfield-type

neural networks to escape from local minima and get superior solutions.

In the search space of a complex-valued multilayer perceptron having

J

hidden units, C-MLP(

J

), there exist flat areas called singular regions, as is the case with a real-valued MLP. The singular regions cause serious stagnation of learning, preventing usual search methods from finding an excellent solution. However, there exist descending paths from the regions since most points in the regions are saddles. This paper proposes a completely new learning method that does not avoid but makes good use of singular regions to stably and successively find excellent solutions commensurate with C-MLP(

J

). Our experiments showed the proposed method worked well.

One usually tries to raise the efficiency of optimization techniques by changing the dynamics of local optimization. In contrast to the above approach, we propose changing the surface of the problem rather than the dynamics of local search. The Mix-Matrix algorithm proposed by the authors previously [1] realizes such transformation and can be applied directly to a max-cut problem and successfully compete with other popular algorithms in this field such as CirCut and Scatter Search.

Model complexities of networks representing multivariable functions is studied in terms of variational norms tailored to types of network units. It is shown that the size of the variational norm reflects both the number of hidden units and sizes of output weights. Lower bounds on growth of variational norms with increasing input dimension

d

are derived for Gaussian units and perceptrons. It is proven that variation of the

d

-dimensional parity with respect to Gaussian Support Vector Machines grows exponentially with

d

and for large values of

d

, almost any randomly-chosen Boolean function has variation with respect to perceptrons depending on

d

exponentially.

In this study we propose a hierarchical neural network that is able to generate a topographical map in its internal layer. The map significantly differs from the conventional Kohonen’s SOM, in that it preserves the topological characteristics in relevance to the context, for example the labels, of the data. This map is useful if we are interested in visualizing the underlying characteristics of the classificability of the data that traditionally cannot be visualized with the standard SOM. In this paper, we expand our network into a multilayered structure that allows us visualize and thus better understand on how the neural network perceives the given data in the light of classification task.

Deep convolutional neural networks are currently applied to computer vision tasks, especially object detection. Due to the large dimensionality of the output space, four dimensions per bounding box of an object, classification techniques do not apply easily. We propose to adapt a structured loss function for neural network training which directly maximizes overlap of the prediction with ground truth bounding boxes. We show how this structured loss can be implemented efficiently, and demonstrate bounding box prediction on two of the Pascal VOC 2007 classes.

Natural communication between humans involves hand gestures, which has an impact on research in human-robot interaction. In a real-world scenario, understanding human gestures by a robot is hard due to several challenges like hand segmentation. To recognize hand postures this paper proposes a novel convolutional implementation. The model is able to recognize hand postures recorded by a robot camera in real-time, in a real-world application scenario. The proposed model was also evaluated with a benchmark database and showed better results than the ones reported in the benchmark paper.

Despite extensive efforts, state-of-the-art detection approaches show a strong degradation of performance with increasing level of occlusion. In this paper we investigate a strategy to improve the detection of occluded objects based on the analytic feature framework from [11] and compare the results in a car detection task. Motivated by an analysis of annotated traffic scenes we focus on a general concept to handle vertical occlusion patterns. For this we describe a two stage classifier architecture that detects vertical car parts in the first stage and combines the local responses in the second. As an extension we provide depth information for the individual car parts helping the classifier in the second stage to reason about typical occlusion patterns.

This paper exploits how Bayesian learning of restricted Boltzmann machine (RBM) can discover more biologically-resembled early visual features. The study is mainly motivated by the sparsity and selectivity of visual neurons’ activations in V1 area. Most previous work of computational modeling emphasize selectivity and sparsity independently, which neglects the underlying connections between them. In this paper, a prior on parameters is defined to simultaneously enhance these two properties, and a Bayesian learning framework of RBM is introduced to infer the maximum posterior of the parameters. The proposed prior performs as the lateral inhibition between neurons. According to our empirical results, the visual features learned from the proposed Bayesian framework yield better discriminative and generalization capability than the ones learned with maximum likelihood, or other state-of-the-art training strategies.

We propose the

Temporal Correlation Net (TCN)

as an object recognition system implementing three basic principles: forming temporal groups of features, learning in a hierarchical structure, and using feedback to predict future input. It is a further development of the Temporal Correlation Graph [1] and shows improved performance on standard datasets like ETH80, COIL100, and ALOI. In contrast to its predecessor it can be trained online on all levels rather than in a level per level batch mode. Training images are presented in temporal order showing objects undergoing specific transformations under viewing conditions the system is supposed to learn invariance under. Computation time and memory demands are low because of sparse learned connectivity and efficient handling of neural activities.

This paper describes an improvement to the Cellular Associative Neural Network, an architecture based on the distributed model of a cellular automaton, allowing it to perform scale invariant pattern matching. The use of tensor products and superposition of patterns allows the system to recall patterns at multiple resolutions simultaneously. Our experimental results show that the architecture is capable of scale invariant pattern matching, but that further investigation is needed to reduce the distortion introduced by image scaling.

Recovering shape from shading is an important problem in computer vision and robotics and many studies have been done. We have already proposed a versatile method of solving the problem by model inclusive learning of neural networks. The method is versatile in the sense that it can solve the problem in various circumstances. Almost all of the methods of recovering shape from shading proposed so far assume that surface reflection properties of a target object are known a priori. It is, however, very difficult to obtain those properties exactly. In this paper we propose a method to resolve this problem by extending our previous method. The proposed method makes it possible to recover shape with simultaneously estimating reflection parameters of an object.

We present a method for biologically-inspired object recognition with one-shot learning of object appearance. We use a computationally efficient model of V1 keypoints to select object parts with the highest information content and model their surroundings using simple colour features. This map-like representation is fed into a dynamical neural network which performs pose, scale and translation estimation of the object given a set of previously observed object views. We demonstrate the feasibility of our algorithm for cognitive robotic scenarios and evaluate classification performance on a dataset of household items.

Visual attention can support object recognition by selecting the relevant target information in the huge amount of sensory data, especially important in scenes composed of multiple objects. Here we demonstrate how attention in a biologically plausible and neuro-computational model of visual perception facilitates object recognition in a robotic real world scenario. We will point out that it is not only important to select the target information, but rather to explicitly suppress the distracting sensory data. We found that suppressing the features of each distractor is not sufficient to achieve robust recognition. Instead, we also have to suppress the location of each distractor. To demonstrate the effect of this spatial suppression, we disable this property and show that the recognition accuracy drops. By this, we show the interplay between attention and suppression in a real world object recognition task.

Visual attention has recently been reported to modulate neural activity of narrow spiking and broad spiking neurons in V4, with increased firing rate and less inter-trial variations. We simulated these physiological phenomena using a neural network model based on spontaneous activity, assuming that the visual attention modulation could be achieved by a change in variance of input firing rate distributed with a lognormal distribution. Consistent with the physiological studies, an increase in firing rate and a decrease in inter-trial variance was simultaneously obtained in the simulation by increasing variance of input firing rate distribution. These results indicate that visual attention forms strong sparse and weak dense input or a ‘winner-take-all’ state, to improve the signal-to-noise ratio of the target information.

A dynamic method of selecting a pruned ensemble of predictors for regression problems is described. The proposed method enhances the prediction accuracy and generalization ability of pruning methods that change the order in which ensemble members are combined. Ordering heuristics attempt to combine accurate yet complementary regressors. The proposed method enhances the performance by modifying the order of aggregation through distributing the regressor selection over the entire dataset. This paper compares four static ensemble pruning approaches with the proposed dynamic method. The experimental comparison is made using MLP regressors on benchmark datasets and on an industrial application of radio frequency source calibration.

This work presents two rejection options. The global rejection option separates the foreign observations – not defined in the classification task – from the normal observations. The local rejection option works after the classification process and separates observations individually for each class. We present implementation of both methods for binary classifiers grouped in a graph structure (tree or directed acyclic graph). Next, we prove that the quality of rejection is identical for both options and depends only on the quality of binary classifiers. The methods are compared on the handwritten digits recognition task. The local rejection option works better for the most part.

In this paper, we first compare the accuracies of the Recursive-Rule Extraction algorithm family, i.e., the Re-RX algorithm, its variant and the “Three-MLP Ensemble by the Re-RX algorithm” (shortened to “Three-MLP Ensemble”) using the Re-RX algorithm as a core part for six kinds of two-class mixed (i.e., discrete and continuous attributes) datasets. Two-class mixed datasets are commonly used for credit scoring and generally in financial domains. In this paper, we compare the accuracy by only the Re-RX algorithm family because of recent comparison reviews and benchmarking study results, obtained by complicated statistics, support vector machines, neuro-fuzzy hybrid classifications, and similar techniques. The Three-MLP Ensemble algorithm cascades standard backpropagation (BP) to train a three neural-network ensemble, where each neural network is a Multi-Layer Perceptron (MLP). Thus, strictly speaking, three neural networks do not need to be trained simultaneously. In addition, the Three-MLP Ensemble is a simple and new concept of rule extraction from neural network ensembles and can avoid previous complicated neural network ensemble structures and the difficulties of rule extraction algorithms. The extremely high accuracy of the Three-MLP Ensemble algorithm generally outperformed the Re-RX algorithm and the variant. The results confirm that the output from the network ensemble can be expressed in the form of rules, and thus opens the “black box” of trained neural network ensembles.

In online gradient descent learning, the local property of the derivative of the output function can cause slow convergence. This phenomenon, called a

plateau

, occurs in the learning process of a multilayer network. Improving the derivative term, we propose a simple method replacing the derivative term with a truncated Gaussian function that greatly increases the convergence speed. We then analyze a soft committee machine trained by proposed method, and show how proposed method breaks a plateau. Results showed that the proposed method eventually led to break the symmetry between hidden units.

Sensitivity analysis became an acknowledged tool used to study the performance of artificial neural networks. Sensitivity analysis allows to assess the influence, e.g., of each neuron or weight on the final network output. In particular various feature selection and pruning strategies are based on this capability. In this paper, we will present a new approximative sensitivity-based training algorithm yielding robust neural networks with generalization capabilities comparable to its exact analytical counterpart, yet much faster.

We present an approach for learning an anisotropic RBF kernel in a game theoretical setting where the value of the game is the degree of separation between positive and negative training examples. The method extends a previously proposed method (KOMD) to perform feature re-weighting and distance metric learning in a kernel-based classification setting. Experiments on several benchmark datasets demonstrate that our method generally outperforms state-of-the-art distance metric learning methods, including the Large Margin Nearest Neighbor Classification family of methods.

We present a framework to address the imbalanced data problem using semi-supervised learning. Specifically, from a supervised problem, we create a semi-supervised problem and then use a semi-supervised learning method to identify the most relevant instances to establish a well-defined training set. We present extensive experimental results, which demonstrate that the proposed framework significantly outperforms all other sampling algorithms in 67% of the cases across three different classifiers and ranks second best for the remaining 33% of the cases.

The RBF network is commonly used for classification and function approximation. The center and radius of the activation function of neurons is an important parameter to be found before the network training. This paper presents a method based on computational geometry to find these coefficients without any parameters provided by the user. The method is compared with a SVM and experimental results showed that our approach is promising.

A new sampling learning method for neural networks is proposed. Derived from an integral representation of neural networks, an

oracle

probability distribution of hidden parameters is introduced. In general rigorous sampling from the oracle distribution holds numerical difficulty, a linear-time sampling algorithm is also developed. Numerical experiments showed that when hidden parameters were initialized by the oracle distribution, following backpropagation converged faster to better parameters than when parameters were initialized by a normal distribution.

In selecting input variables by block addition and block deletion (BABD), multiple input variables are added and then deleted, keeping the cross-validation error below that using all the input variables. The major problem of this method is that selection time becomes large as the number of input variables increases. To alleviate this problem, in this paper, we propose incremental block addition and block deletion of input variables. In this method, for an initial subset of input variables we select input variables by BABD. Then in the incremental step, we add some input variables that are not added before to the current selected input variables and iterate BABD. To guarantee that the cross-validation error decreases monotonically by incremental BABD, we undo incremental BABD if the obtained cross-validation error rate is worse than that at the previous incremental step. We evaluate incremental BABD using some benchmark data sets and show that by incremental BABD, input variable selection is speeded up with the approximation error comparable to that by batch BABD.

Recent work in supervised learning has shown that naive Bayes text classifiers with strong assumptions of independence among features, such as multinomial naive Bayes (MNB), complement naive Bayes (CNB) and the one-versus-all-but-one model (OVA), have achieved remarkable classification performance. This fact raises the question of whether a naive Bayes text classifier with less restrictive assumptions can perform even better. Responding to this question, we firstly evaluate the correlation-based feature selection (CFS) approach in this paper and find that it performs even worse than the original versions. Then, we propose a CFS-based feature weighting approach to these naive Bayes text classifiers. We call our feature weighted versions FWMNB, FWCNB and FWOVA respectively. Our proposed approach weakens the strong assumptions of independence among features by weighting the correlated features. The experimental results on a large suite of benchmark datasets show that our feature weighted versions significantly outperform the original versions in terms of classification accuracy.

Classification with rejection is well understood for classifiers which provide explicit class probabilities. The situation is more complicated for popular deterministic classifiers such as learning vector quantisation schemes: albeit reject options using simple distance-based geometric measures were proposed [4], their local scaling behaviour is unclear for complex problems. Here, we propose a local threshold selection strategy which automatically adjusts suitable threshold values for reject options in prototype-based classifiers from given data. We compare this local threshold strategy to a global choice on artificial and benchmark data sets; we show that local thresholds enhance the classification results in comparison to global ones, and they better approximate optimal Bayesian rejection in cases where the latter is available.

More complex data formats and dedicated structure metrics have spurred the development of intuitive machine learning techniques which directly deal with dissimilarity data, such as relational learning vector quantization (RLVQ). The adjustment of metric parameters like relevance weights for basic structural elements constitutes a crucial issue therein, and first methods to automatically learn metric parameters from given data were proposed recently. In this contribution, we investigate a robust learning scheme to adapt metric parameters such as the scoring matrix in sequence alignment in conjunction with prototype learning, and we investigate the suitability of efficient approximations thereof.

The Adaline network [1] is a classic neural architecture whose learning rule is the famous least mean squares (LMS) algorithm (a.k.a. delta rule or Widrow-Hoff rule). It has been demonstrated that the LMS algorithm is optimal in

H

 ∞ 

sense since it tolerates

small

(in energy) disturbances, such as measurement noise, parameter drifting and modelling errors [2,3]. Such optimality of the LMS algorithm, however, has been demonstrated for regression-like problems only, not for pattern classification. Bearing this in mind, we firstly show that the performances of the LMS algorithm and variants of it (including the recent Kernel LMS algorithm) in pattern classification tasks deteriorates considerably in the presence of labelling errors, and then introduce robust extensions of the Adaline network that can deal efficiently with such errors. Comprehensive computer simulations show that the proposed extension consistently outperforms the original version.

Support Vector Machines (SVMs) have been recognized as one of the most successful classification methods for many applications in static environment. However in dynamic environment, data characteristics may evolve over time. This leads to deteriorate dramatically the performance of SVMs over time. This is because of the use of data which is no more consistent with the characteristics of new incoming one. Thus in this paper, we propose an approach to recognize and handle concept changes with support vector machine. This approach integrates a mechanism to use only the recent and most representative patterns to update the SVMs without a catastrophic forgetting.

We propose Two Subspace-based Kernel Local Discriminant Embedding (TSKLDE) for feature extraction and recognition. The procedure of TSKLDE is divided into two stages. The first stage applies the Kernel Principal Component Analysis (KPCA) to the raw data. Based on the output of KPCA, the second stage seeks two kinds of discriminant information, regular and irregular. These are based on transforms derived from the within-class locality preserving scatter. The resulting framework retain the advantages of supervised global and local techniques. Besides, it is an inductive nonlinear technique in the sense that novel examples or images can be directly mapped. The proposed algorithm was tested and evaluated using three public face databases. The experimental results show that the proposed TSKLDE outperforms other Kernel based algorithms.

We study how excitation and inhibition are distributed mesoscopically in small brain regions, by means of a computational model of coupled cortical columns described by neural mass models. Two cortical columns coupled bidirectionally through both excitatory and inhibitory connections can spontaneously organize in a regime in which one of the columns is purely excitatory and the other is purely inhibitory, provided the excitatory and inhibitory coupling strengths are adequately tuned. We also study the case of three columns in different coupling configurations (linear array and all-to-all coupling), finding abrupt transitions between heterogeneous and homogeneous excitatory/inhibitory patterns and strong multistability in their distribution.

Strong experimental evidence suggests that cortical memory traces are consolidated during off-line memory reprocessing that occurs in the off-line states of sleep or waking rest. It is unclear, what plasticity mechanisms are involved in this process and what changes are induced in the network in the off-line regime. Here, we examine a hierarchical recurrent neural network that performs unsupervised learning on natural face images of different persons. The proposed network is able to self-generate memory replay while it is decoupled from external stimuli. Remarkably, the recognition performance is tremendously boosted after this off-line regime specifically for the novel face views that were not shown during the initial learning. This effect is independent of synapse-specific plasticity, relying completely on homeostatic regulation of intrinsic excitability. Comparing a purely feed-forward network configuration with the full version reveals a substantially stronger boost in recognition performance for the fully recurrent network architecture after the off-line regime.

LIPNet is a pyramidal neural network with lateral inhibition developed for pattern recognition, inspired in the concept of receptive and inhibitory fields from the human visual system. Although this network can implicitly extract features and use these features to properly classify patterns in images, many parameters must be defined prior to the network training and operation. Besides, these parameters have a huge impact on the recognition performance. This paper proposes an encoding scheme aiming at optimizing the LIPNet structure using Particle Swarm Optimization. Preliminary results for a face detection problem using a well known benchmark set showed that our approach achieved better classification rates when compared to the original LIPNet.

Grid cells in the dorsocaudal medial entorhinal cortex (dMEC) of the rat provide a metric representation of the animal’s local environment. The collective firing patterns in a network of grid cells forms a triangular mesh that accurately tracks the location of the animal. The activity of a grid cell network, similar to head direction cells, displays path integration characteristics. Classical robotics use path integrators in the form of inertial navigation systems to track spatial information of an agent as well. In this paper, we describe an implementation of a network of grid cells as a dead reckoning system for the PR2 robot.

Brains experience sensory information grounded in sensor- relative frames of reference. To compare sensory information from different sensor sources, such as vision and touch, this information needs to be mapped onto each other. To do so, the brain needs to learn suitable spatial transformations and the literature suggests that gain fields accomplish such transformations. However, when transforming three dimensional spaces or even six dimensional configuration spaces then simple gain fields do not scale to such a dimensionality. We are investigating how this curse of dimensionality can be overcome. Based on neural population-encoded, component-wise spatial representations, we show that a hierarchy of gain fields can accomplish higher-dimensional transformations and that its weights can be learned effectively by means of standard backpropagation.

One of the key computations performed in human brain is multi-sensory cue integration, through which humans are able to estimate the current state of the world to discover relative reliabilities and relations between observed cues. Mammalian cortex consists of highly distributed and interconnected populations of neurons, each providing a specific type of information about the state of the world. Connections between areas seemingly realize functional relationships amongst them and computation occurs by each area trying to be consistent with the areas it is connected to. In this paper using line-attraction dynamics and divisive normalization, we present a computational framework which is able to learn arbitrary non-linear relations between multiple cues using a simple Hebbian Learning principle. After learning, the network dynamics converges to the stable state so to satisfy the relation between connected populations. This network can perform several principle computational tasks such as inference, de-noising and cue-integration. By applying a real world multi-sensory integrating scenario, we demonstrate that the network can encode relative reliabilities of cues in different areas of the state space, over distributed population vectors. This reliability based encoding biases the network’s dynamics in favor of more reliable cues and realizes a near optimal sensory integration mechanism. Additional important features of the network are its scalability to cases with higher order of modalities and its flexibility to learn smooth functions of relations which is necessary for a system to operate in a dynamic environment.

Looking is one of the most basic and fundamental goal-directed behaviors. The neural circuitry that generates gaze shifts towards target objects is adaptive and compensates for changes in the sensorimotor plant. Here, we present a neural-dynamic architecture, which enables an embodied agent to direct its gaze towards salient objects in its environment. The sensorimotor mapping, which is needed to accurately plan the gaze shifts, is initially learned and is constantly updated by a gain adaptation mechanism. We implemented the architecture in a simulated robotic agent and demonstrated autonomous map learning and adaptation in an embodied setting.

The visual recognition of body motion in the primate brain requires the temporal integration of information over complex patterns, potentially exploiting recurrent neural networks consisting of shape- and optic-flow-selective neurons. The paper presents a mathematically simple neurodynamical model that approximates the mean-field dynamics of such networks. It is based on a two-dimensional neural field with appropriate lateral interaction kernel and an adaptation process for the individual neurons. The model accounts for a number of, so far not modeled, observations in the recognition of body motion, including perceptual multi-stability and the weakness of repetition suppression, as observed in single-cell recordings for the repeated presentation of action stimuli. In addition, the model predicts novel effects in the perceptual organization of action stimuli.

In this article, we present an original neural space/latency code, integrated in a multi-layered neural hierarchy, that offers a new perspective on probabilistic inference operations. Our work is based on the dynamic neural field paradigm that leads to the emergence of activity bumps, based on recurrent lateral interactions, thus providing a spatial coding of information. We propose that lateral connections represent a data model, i.e., the conditional probability of a “true” stimulus given a noisy input. We propose furthermore that the resulting attractor state encodes the most likely ”true” stimulus given the data model, and that its latency expresses the confidence in this interpretation. Thus, the main feature of this network is its ability to represent, transmit and integrate probabilistic information at multiple levels so that to take near-optimal decisions when inputs are contradictory, noisy or missing. We illustrate these properties on a three-layered neural hierarchy receiving inputs from a simplified robotic object recognition task. We also compare the network dynamics to an explicit probabilistic model of the task, to verify that it indeed reproduces all relevant properties of probabilistic processing.

Several computational models have been designed to help our understanding of the conditions under which persistent activity can be sustained in cortical circuits during working memory tasks. Here we focus on one such model that has shown promise, that uses polychronization and short term synaptic dynamics to achieve this reverberation, and explore it with respect to different physiological parameters in the brain, including size of the network, number of synaptic connections, small-world connectivity, maximum axonal conduction delays, and type of cells (excitatory or inhibitory). We show that excitation and axonal conduction delays greatly affect the sustainability of spatio-temporal patterns of spikes called polychronous groups.

Brain functions such as learning and memory rely on synaptic plasticity. Many studies have shown that synaptic plasticity can be driven by the timings between pre- and postsynaptic spikes, also known as spike-timing-dependent plasticity (STDP). In most of the modeling studies exploring STDP functions, presynaptic spikes have been postulated to be Poisson (random) spikes and postsynaptic neurons have been described using an integrate-and-fire model, for simplicity. However, experimental data suggest this is not necessarily true. In this study, we investigated how STDP worked in synaptic competition if more neurophysiologically realistic properties for pre- and postsynaptic dynamics were incorporated; presynaptic (input) spikes obeyed a gamma process and the postsynaptic neuron was a multi-timescale adaptive threshold model. Our results showed that STDP strengthened specific combinations of pre- and postsynaptic properties; regular spiking neurons favored regular input spikes whereas random spiking neurons did random ones, suggesting neural information coding utilizes both the properties.

The striatum comprises part of a feedback loop between the cerebral cortex, thalamus and other nuclei of the basal ganglia, ultimately guiding action selection and motor learning. Much of this is facilitated by striatal projection neurons, which receive and process highly convergent cortical and thalamic excitatory inputs. All of the glutamatergic inputs to projection neurons synapse on dendrites, many directly on spine heads. The distal, but not proximal, dendrites of projection neurons are capable of supporting synaptically driven regenerative events, which are transfered to the soma as depolarized upstates from which action potentials can occur. In this study we present a modified NEURON model of a striatal projection neuron, and use it to examine the location-dependence of upstate generation and action potential gating. Specifically, simulations show that the small diameter of distal SPN dendrites can support plateau potentials by increasing the cooperativity among neighboring spines. Furthermore, such distally evoked plateaus can boost the somatic response to stimulation of proximal dendritic spines, facilitating action potential generation. The implications these results have for action selection are discussed.

Reward-modulated learning rules for spiking neural networks have emerged, that have been demonstrated to solve a wide range of reinforcement learning tasks. Despite this, little work has aimed to classify spike patterns by the timing of output spikes. Here, we apply a rewardmaximising learning rule to teach a spiking neural network to classify input patterns by the latency of output spikes. Furthermore, we compare the performance of two escape rate functions that drive output spiking activity: the Arrhenius & Current (A&C) model and Exponential (EXP) model. We find A&C consistently outperforms EXP, and especially in terms of the time taken to converge in learning. We also show that jittering input patterns with a low noise amplitude leads to an improvement in learning, by reducing the variation in the performance.

Surveys are widely conducted as a means to obtain information on thoughts, opinions and feelings of people. The representativeness of a sample is a major concern in using surveys. In this article, we consider meaning variation which is another potentially remarkable but less studied source of problems. We use Grounded Intersubjective Concept Analysis (GICA) method to quantify meaning variation and demonstrate the effect on survey analysis through a case study in which food prices and food concepts are considered.

In this paper, we extract various patterns of the spatio-temporal distribution from Foursquare. Foursquare is a location-based social networking system which has been widely used recently. For extracting patterns, we employ ICA (Independent Component Analysis), which is a useful method in signal processing and feature extraction. Because the Foursquare dataset consists of check-in’s of users at some time points and locations, ICA is not directly applicable to it. In order to smooth the dataset, we estimate a continuous spatio-temporal distribution by employing a diffusion-type formula. The experiments on an actual Foursquare dataset showed that the proposed method could extract some plausible and interesting spatio-temporal patterns.

Collaborative filtering and Content-based filtering methods are two famous methods used by recommender systems. Restricted Boltzmann Machine(RBM) model rivals the best collaborative filtering methods, but it focuses on modeling the correlation between item ratings. In this paper, we extend RBM model by incorporating content-based features such as user demograohic information, items categorization and other features. We use Naive Bayes classifier to approximate the missing entries in the user-item rating matrix, and then apply the modified UI-RBM on the denser rating matrix. We present expermental results that show how our approach, Content-boosted Restricted Boltzmann Machine(CB-RBM), performs better than a pure RBM model and other content-boosted collaborative filtering methods.

This paper introduces Financial Self–Organizing Maps (FinSOM) as a SOM sub–class where the mapping of inputs on the neural space takes place using functions with economic soundness, that makes them particularly well–suited to analyze financial data. The visualization capabilities as well as the explicative power of both the standard SOM and the FinSOM variants is tested on data from the German Stock Exchange. The results suggest that, dealing with financial data, the FinSOM seem to offer superior representation capabilities of the observed phenomena.

Mapping, localization and navigation are major topics and challenges for mobile robotics. To perform tasks and to interact efficiently in the environment, a robot needs knowledge about its surroundings. Many robots today are capable of performing simultaneous mapping and localization to generate own world representations. Most assume an array of highly sophisticated artificial sensors to track landmarks placed in the environment. Recently, there has been significant interest in research approaches inspired by nature and RatSLAM is one of them. It has been introduced and tested on wheeled robots with good results. To examine how RatSLAM behaves on humanoid robots, we adapt this model for the first time to this platform by adjusting the given constraints. Furthermore, we introduce a multiple hypotheses mapping technique which improves mapping robustness in open spaces with features visible from several distant locations.

In this work, we propose the use of support vector regression ensembles for wind power prediction. Ensemble methods often yield better classification and regression accuracy than classical machine learning algorithms and reduce the computational cost. In the field of wind power generation, the integration into the smart grid is only possible with a precise forecast computed in a reasonable time. Our support vector regression ensemble approach uses bootstrap aggregating (bagging), which can easily be parallelized. A set of weak predictors is trained and then combined to an ensemble by aggregating the predictions. We investigate how to choose and train the individual predictors and how to weight them best in the prediction ensemble. In a comprehensive experimental analysis, we show that our SVR ensemble approach renders significantly better forecast results than state-of-the-art predictors.

The studied inverse problem is determination of partial concentrations of inorganic salts in multi-component water solutions by their Raman spectra. The problem is naturally divided into two parts: 1) determination of the component composition of the solution, i.e. which salts are present and which not; 2) determination of the partial concentration of each of the salts present in the solution. Within the first approach, both parts of the problem are solved simultaneously, with a single neural network (perceptron) with several outputs, each of them estimating the concentration of the corresponding salt. The second approach uses data clusterization by Kohonen networks for consequent identification of component composition of the solution by the cluster, which the spectrum of this solution falls into. Both approaches and their results are discussed in this paper.

Optical distortions of an image are created in astronomical, optical microscopy and communication systems by light propagation throughout a variety of optical components. Usually, optical aberrations are corrected by using an Adaptive Optics system, where a wavefront sensor is used to measure the optical distortion. In this work, we propose to use a Lateral Inhibition Pyramidal Neural Network (LIPNet) in the frequency domain to classify optical defocus (using Zernike polynomials

Z

5

), such that optical defocus value can be detected directly from the image without the use of wavefront sensing. The results show the potentiality of the method and open new opportunities to explore this kind of neural networks algorithms for wavefront sensing and Adaptive optic systems.

The SpiNNaker neural computing project has created a hardware architecture capable of scaling up to a system with more than a million embedded cores, in order to simulate more than one billion spiking neurons in biological real time. The heart of this system is the SpiNNaker chip, a multi-processor System-on-Chip with a high level of interconnectivity between its processing units. Here we present a Dynamically Extendable SpiNNaker Chip Computing Module that allows a SpiNNaker machine to be deployed on small mobile robots. A non-neural application, the simulation of the movement of a flock of birds, was developed to demonstrate the general purpose capabilities of this new platform. The developed SpiNNaker machine allows the simulation of up to one million spiking neurons in real time with a single SpiNNaker chip and is scalable up to 256 computing nodes in its current state.

Recently a new technique called multiband imaging was introduced, it allows extremely low repetition times for functional magnetic resonance imaging (fMRI). As these ultra fast imaging scans can increase the Nyquist rate by an order of magnitude, there are many new effects, that have to be accounted for. As more frequencies can now be sampled directly, we want to analyze especially those that are due to physiological noise, such as cardiac and respiratory signals. Here, we adapted RETROICOR [4] to handle multiband fMRI data. We show the importance of physiological noise regression for standard temporal resolution fMRI and compare it to the high temporal resolution case. Our results show that especially for multiband fMRI scans, it is of the utmost importance to apply physiological noise regression, as residuals of these noises are clearly detectable in non noise independent components if no prior physiological noise has been applied.

Melanoma is the most deathly of all skin cancers but early diagnosis can ensure a high degree of survival. Early diagnosis is one of the greatest challenges due to lack of experience of general practitioners (GPs). In this paper we present a clinical decision support system designed for general practitioners, aimed at saving time and resources in the diagnostic process. Segmentation, pattern recognition, and change detection are the important steps in our approach. This paper also investigates the performance of Artificial Neural Network (ANN) learning algorithms for skin cancer diagnosis. The capabilities of three learning algorithms i.e. Levenberg-Marquardt (LM), Resilient Back propagation (RP), Scaled Conjugate Gradient (SCG) algorithms in differentiating melanoma and benign lesions are studied and their performances are compared. The results show that Levenberg-Marquardt algorithm was quick and efficient in figuring out benign lesions with specificity 95.1%, while SCG algorithm gave better results in detecting melanoma at the cost of more number of epochs with sensitivity 92.6%.