Artificial Neural Networks and Machine Learning – ICANN 2012

The Artificial Reaction Network (ARN) is a bio-inspired connectionist paradigm based on the emerging field of Cellular Intelligence. It has properties in common with both AI and Systems Biology techniques including Artificial Neural Networks, Petri Nets, and S-Systems. This paper discusses the temporal aspects of the ARN model using robotic gaits as an example and compares it with properties of Artificial Neural Networks. The comparison shows that the ARN based network has similar functionality.

The standard Hopfield model is generalized to the case when input patterns are provided with weights that are proportional to the frequencies of patterns occurrence at the learning process. The main equation is derived by methods of statistical physics, and is solved for an arbitrary distribution of weights. An infinitely large number of input patterns can be written down in connection matrix however the memory of the network will consist of patterns whose weights exceed a critical value. The approach eliminates the catastrophic destruction of the memory characteristic to the standard Hopfield model.

In this paper, we investigate the dynamics and the global exponential stability of the Hopfield Neural network with time-varying delay and variable coefficients. For this purpose, the activation functions are assumed to be globally Lipschitz continuous. The properties of norms and the contraction principle are adjusted to ensure the existence as well as the uniqueness of the the pseudo almost automorphic solution. Then by employing suitable analytic techniques, global attractivity of the unique pseudo almost automorphic solution is established.

A dilemma arises when sequence generation is implemented on embodied autonomous agents. While achieving an individual action goal, the agent must be in a stable state to link to fluctuating and time-varying sensory inputs. To transition to the next goal, the previous state must be released from stability. A previous proposal of a neural dynamics solved this dilemma by inducing an instability when a “condition of satisfaction” signals that an action goal has been reached. The required structure of dynamic coupling limited the complexity and flexibility of sequence generation, however. We address this limitation by showing how the neural dynamics can be generalized to generate hierarchically structured behaviors. Directed couplings downward in the hierarchy initiate chunks of actions, directed couplings upward in the hierarchy signal their termination. We analyze the mathematical mechanisms and demonstrate the flexibility of the scheme in simulation.

The problem of learning Bayesian network structure is well known to be NP–hard. It is therefore very important to develop efficient approximation techniques. We introduce an algorithm that within the framework of influence diagrams translates the structure learning problem into the strategy optimisation problem, for which we apply the Chen’s self–annealing stochastic optimisation algorithm. The effectiveness of our method has been tested on computer–generated examples.

In the paper we deal with the NP-complete problem of minimization a quadratic form of

N

binary variables. The minimization approach based on extensive random search is considered. To increase the efficiency of the random-search algorithm, we vary the attraction area of the deepest minima of the functional by changing the matrix

T

it is based on. The new matrix

M

, called

mix-matrix

, is a mixture of

T

and

T

2

. We demonstrate that such a substitution brings about changes of the energy surface: deep minima displace very slightly in the space (the Hemming distance of the shift is of about 0.01*

N

), they become still deeper and their attraction areas grow significantly. At the same time the probability of finding close to optimal solutions increases abruptly (by 2-3 orders of magnitude in case of a 2D Ising model of size 12

×

12 and in case of dense instances of size 500).

This paper describes an architecture based on superimposed distributed representations and distributed associative memories which is capable of performing rule chaining. The use of a distributed representation allows the system to utilise memory efficiently, and the use of superposition reduces the time complexity of a tree search to

O

(

d

), where

d

is the depth of the tree. Our experimental results show that the architecture is capable of rule chaining effectively, but that further investigation is needed to address capacity considerations.

The paper introduces a novel approach for defining efficient generative kernels for structured-data based on the concept of multisets and Jaccard similarity. The multiset feature-space allows to enhance the adaptive kernel with syntactic information on structure matching. The proposed approach is validated using an input-driven hidden Markov model for trees as generative model, but it is enough general to be straightforwardly applicable to any probabilistic latent variable model. The experimental evaluation shows that the proposed Jaccard kernel has a superior classification performance with respect to the Fisher Kernel, while consistently reducing the computational requirements.

Nonnegative Matrix Factorization (NMF) is an unsupervised learning method that has been already applied to many applications of spectral signal unmixing. However, its efficiency in some applications strongly depends on optimization algorithms used for estimating the underlying nonnegatively constrained subproblems. In this paper, we attempt to tackle the optimization tasks with the inexact Interior-Point (IP) algorithm that has been successfully applied to image deblurring [S. Bonettini, T. Serafini, 2009]. The experiments demonstrate that the proposed NMF algorithm considerably outperforms the well-known NMF algorithms for blind unmixing of the Raman spectra.

This paper introduces a hybrid optimized polynomial neural network (HOPNN), a novel architecture that is constructed by using a combination of fuzzy rule-based models, polynomial neural networks (PNNs) and a hybrid optimization algorithm. The proposed hybrid optimization algorithm is developed by a combination of a space search algorithm and Powell’s method. The structure of this HOPNN comprises of a synergistic usage of fuzzy-relation-based polynomial neurons and polynomial neural network. The fuzzy-relation-based polynomial neurons are fuzzy rule-based models, while the polynomial neural network is an extended group method of data handling (GMDH). The architecture of HOPNN is essentially modified PNNs whose basic nodes of the first (input) layer are fuzzy-rule-based polynomial neurons rather than conventional polynomial neurons. Moreover, the proposed hybrid optimization algorithm is exploited to optimize the structure topology of HOPNN. The performance of the network is quantified through experimentation in which we use a number of modeling benchmarks already experimented with in the realm of fuzzy or neurofuzzy modeling.

In this paper, we study a Tikhonov-type regularization for restricted Boltzmann machines (RBM). We present two alternative formulations of the Tikhonov-type regularization which encourage an RBM to learn a smoother probability distribution. Both formulations turn out to be combinations of the widely used weight-decay and sparsity regularization. We empirically evaluate the effect of the proposed regularization schemes and show that the use of them could help extracting better discriminative features with sparser hidden activation probabilities.

This paper uses the results of the characterization of amorphous semiconductor thin film transistors (TFTs) with a split gate and the quasi-permanent memory structure referred to as silicon oxide nitride semiconductor (SONOS) gates, to model spiking neural circuits with Hebbian learning ability. MOSFETs using organic (tris 8-hydroxyquinolinate aluminum (Alq3), copper phthalocyanine (CuPc)) and inorganic (ZnO) amorphous materials can be fabricated with split gates, which will provide multiple synaptic inputs. A simple Hebbian learning circuit is added to charge and discharge the SONOS device. The primary result of this work is the demonstration of the practicality of using SONOS amorphous organic TFTs with multiple gates and imbedded Hebbian learning capability in spiking neuron analog circuits. The use of these elements allows for the design and fabrication of high-density 3-dimensional circuits that can achieve the interconnect density of biological neural systems.

A FPGA hardware core has been designed for real-time network simulations with up to 400 physiologically realistic, conductance-based neurons of the Hodgkin-Huxley type. A PC-FPGA interface allows easy parameter adjustment and on-line display of basic synchronization measures like field potentials, spike times or color-coded voltages of the complete array. Simulations of 20 ·20 gap-junction coupled 4-dimensional neurons reveal remarkable alterations of the synchronization states and impulse patterns during linearly increasing coupling strengths.

Sodium entry during the decaying phase of the action potential determines the metabolic efficiency of a neuron’s spiking mechanism. Recent studies have reported that mammalian action potentials are close to metabolic optimality but that fast-spiking inhibitory neurons are less efficient than their pyramidal counterparts. It is postulated that this represents nature’s tradeoff between metabolic efficiency and the ability to discharge at high rates. Using eight different published Hodgkin-Huxley models of mammalian neurons to cover a wide range of action potential metabolic efficiencies, we show that the cost of operating a neuron is heavily dependent on its output frequency. We observe that this cost is significantly smaller than the frequency-dependent cost naively estimated from an isolated action potential, the gap increasing with increasing frequencies. Our results demonstrate that metabolic efficiency cannot be considered only in terms of isolated action potentials but must be studied over a range of meaningful frequencies.

Spiking neural networks (SNN) aim to mimic membrane potential dynamics of biological neurons. They have been used widely in neuromorphic applications and neuroscience modeling studies. We design a parallel SNN accelerator for producing large-scale cortical simulation targeting an off-the-shelf Field-Programmable Gate Array (FPGA)-based system. The accelerator parallelizes synaptic processing with run time proportional to the firing rate of the network. Using only one FPGA, this accelerator is estimated to support simulation of 64K neurons 2.5 times real-time, and achieves a spike delivery rate which is at least 1.4 times faster than a recent GPU accelerator with a benchmark toroidal network.

We use neuromorphic chips to perform arbitrary mathematical computations for the first time. Static and dynamic computations are realized with heterogeneous spiking silicon neurons by programming their weighted connections. Using 4K neurons with 16M feed-forward or recurrent synaptic connections, formed by 256K local arbors, we communicate a scalar stimulus, quadratically transform its value, and compute its time integral. Our approach provides a promising alternative for extremely power-constrained embedded controllers, such as fully implantable neuroprosthetic decoders.

This paper presents the SEPELYNS architecture that permits to interconnect multiple spiking neurons focused on hardware implementations. SEPELYNS can connect millions of neurons with thousands of synapses per neuron in a layered fabric that provides some capabilities such as connectivity, expansion, flexibility, bio-plausibility and reusing of resources that allows simulation of very large networks. We present the three layers of this architecture (neuronal, network adapters and networks on chip layers) and explain its performance parameters such as throughput, latency and hardware resources. Some application examples of large neural networks on SEPELYNS are studied; these will show that use of on-chip parallel networks could permit the hardware simulation of populations of spiking neurons.

We propose an associative learning model using reward modulated spike-time dependent plasticity in reinforcement learning paradigm. The task of learning is to associate a stimulus pair, known as the

predictor

 − 

choice

pair, to a target response. In our model, a generic architecture of neural network has been used, with minimal assumption about the network dynamics. We demonstrate that stimulus-stimulus-response association can be implemented in a stochastic way within a noisy setting. The network has rich dynamics resulting from its recurrent connectivity and background activity. The algorithm can learn temporal sequence detection and solve temporal XOR problem.

A synaptic architecture featuring both excitatory and inhibitory neurons is assembled aiming to build up an associative memory system. The connections follow a hebbian-like rule. The network activity is analyzed using a multidimensional reduction method, Principal Component Analysis (PCA), applied to neuron firing rates. The patterns are discriminated and recognized by well defined paths that emerge within PCA subspaces, one for each pattern. Detailed comparisons among these subspaces are used to evaluate the network storage capacity. We show a transition from a retrieval to a non-retrieval regime as the number of stored patterns increases. When gap junctions are implemented together with the chemical synapses, this transition is shifted and a larger number of memories is associated to the network.

This paper studies the bifurcating neuron whose base signal is given by filtering a source signal. As the time constant of the filter varies, shape of base signal varies and the neuron can exhibit rich phenomena. We show typical ones: co-existing periodic and chaotic orbits in single neuron; and chaos + chaos = order and order + order = chaos in the pulse-coupled neurons. The phenomena can be analyzed by the one-dimensional map of spike-phases. Presenting a simple test circuit, typical phenomena are confirmed experimentally.

This paper studies digital spike maps that can generate various periodic spike-trains. In order to analyze the maps, we present a simple analysis algorithm to calculate basic feature quantities. We then analyze a typical example of the map given by discretizing the bifurcating neuron. Applying the algorithm to the example, we demonstrate complex dynamics and give basic classification of the dynamics.

In this work a novel approach to automatically generate hardware is introduced that allows accelerated simulation of artificial neural networks (ANN) on field-programming gate arrays (FPGAs). A compiler architecture has been designed that primarily aims at reducing the development effort for non-hardware developers. This is done by implementing automatic generation of accordingly adjusted hardware processors. Deduced from high-level OpenCL source code, the processors are able to spatially map ANNs in a massive parallel fashion.

Basal ganglia circuits are known to have role in a wide range of behaviour spanning from movement initiation to high order cognitive processes as reward related learning. Here, the intention is to have a biophysically realistic model of basal ganglia circuit for voluntary motor action selection. The ultimate aim is to provide a framework for models which could help comprehension of complex processes. To fulfill this aim a model capable of simulating direct, indirect and hyperdirect pathways with modified Hodgkin-Huxley neuron model is proposed. This model considers more neural structures than the works similar in the literature and can simulate activity of neurons in the neural structures taking part in action selection. The model proposed is shown to be versatile as the simulation results obtained are similar to the neuron activity recordings of the considered neural structures published previously.

Huntington’s is a genetic, progressive neuro-degenerative disease, causing massive network degradation effecting the medium spiny neurons of the striatum in the basal ganglia (a set of sub-cortical nuclei, believed to be critical for action selection). Despite substantial striatal cell atrophy, some cognitive functions have been shown to improve in manifest Huntington’s disease patients over healthy and pre-symptomatic Huntington’s disease patients. Using a detailed model of the striatal microcircuit, we show that combining current ideas about the underlying causes of the disease could lead to the counter-intuitive result of improved competitive network dynamics for signal selection.

Recent experimental evidence shows that cellular connectivity in the brain is

not random

. More specifically, bidirectional connections between pairs of excitatory neurons are predominantly found when neurons connect by short-term facilitating synapses. For this type of synapse, excitatory postsynaptic potentials (EPSPs) transiently increase upon repeated presynaptic activation. Unidirectional connections between pairs of excitatory neurons, however, are predominantly found when neurons are connected by short-term depressing synapses. For these synapses, EPSPs transiently attenuate upon repeated activation. Here we present a simple model that combines Short-Term Plasticity (STP) and Spike-Timing Dependent Plasticity (STDP) that might explain the correlation between specific synaptic dynamics and, unidirectional or bidirectional, connectivity patterns.

The Multi-electrode Array (MEA) technology allows the in-vitro culture of neuronal networks that can be used as a simplified and accessible model of the central nervous system, given that they exhibit activity patterns similar to the in-vivo tissue. Current devices generate huge amounts of data, thus motivating the development of systems capable of discriminating diverse cultured neuronal network activity patterns. In this paper, we describe the use of Inter-Spike Interval statistics coupled to data compression statistics in two discrimination applications. One of them concerning spontaneous vs. stimulated activity patterns, and the other concerning spontaneous responses from a control culture of neurons and a previously treated one. We show that the data compression ratio of the trains of spikes emerging from those cultures can be used to enhance the discrimination performance.

We present a biologically inspired recurrent network of spiking neurons and a learning rule that enables the network to balance a ball on a flat circular arena and to steer it towards a target field, by controlling the inclination angles of the arena. The neural controller is a recurrent network of adaptive exponential integrate and fire neurons configured and connected to match properties of cortical layer IV. The network is used as a liquid state machine with four action cells as readout neurons. The solution of the task requires the controller to take its own reaction time into account by anticipating the future state of the controlled system. We demonstrate that the cortical network can robustly learn this task using a supervised learning rule that penalizes the error on the force applied to the arena.

This article introduces a bio-inspired neural network that uses timing for streaming self-generated signals. Timing messages in a crowded informatics space is a well-known method for collision avoidance (i.e. 802.11 protocol). Recently, Nogueira and Caputi (2011) have shown that timing the emission of a sensory carrier allows electric fish to stream self-generated signals through a refractoriness window in the presence of interference. Here I model the system and show that a simple sensory-motor feedback loop is enough for adapting the timing of the pacemaker controlling the emission of the carrier. Critical aspects of this behavior are the shape of the refractoriness window and the duration of signals effects on the motor command.

The image matching is an important technique in the image processing and the method using Pulse Coupled Neural Network (PCNN) had been proposed. One of the useful feature of the method is that the method is valid for the image matching among rotated, magnified and shrunk images. We have been proposed the parameter learning method of the PCNN for the image matching. Considering that the image matching technique will utilize for any advanced image processing such as a content based image retrieval, the capacity and the versatility of the method are important characteristics to evaluate the method. In this study, our method is tested using total 17,920 images and we describe the characteristics of the capacity and the versatility of image matching method using PCNN with our parameter learning algorithm.

A spiking neuron model described by an asynchronous cellular automaton is introduced. Our model can be implemented in an asynchronous sequential logic circuit and its control parameter is adjustable after implementation in an FPGA. It is shown that our model can reproduce the features of four groups into which biological and other model neurons are classified. In addition, underlying bifurcations of the four groups are analyzed, and the results yield basic guides to synthesis of our model.

This paper presents a spiking neural network (SNN) for binaural sound source localisation (SSL). The cues used for SSL were the interaural time (ITD) and level (ILD) differences. ITDs and ILDs were extracted with models of the medial superior olive (MSO) and the lateral superior olive (LSO). The MSO and LSO outputs were integrated in a model of the inferior colliculus (IC). The connection weights between the MSO and LSO neurons to the IC neurons were estimated using Bayesian inference. This inference process allowed the algorithm to perform robustly on a robot with ~40,dB of ego-noise. The results showed that the algorithm is capable of differentiating sounds with an accuracy of 15°.

Bayesian estimation theory has been expected to explain how brain deals with uncertainty such as feature extraction against noisy observations. It has been implied that the neural networks that model cortical network could implement the Bayesian estimation algorithm by several previous studies. However, it is still unclear whether it is possible to implement the required computational procedures of the algorithm under physiological and anatomical constraints of the neural systems. We here propose the neural network that implements the algorithm in a biologically realizable manner, incorporating the discrete choice theory into the previously proposed model. Our model successfully demonstrated an orientation discrimination task with significantly noisy visual images.

Spiking neural networks (SNNs) are considered to be more biologically realistic compared to typical rate-coded networks as they can model closely different types of neurons and their temporal dynamics. Typical spiking models use a number of fixed parameters such as the ratio between excitatory and inhibitory neurons. However, the parameters that are used in these models focus almost exclusively on our understanding of the neocortex with, for example, 80% of neurons chosen as excitatory and 20% inhibitory. In this paper we will evaluate how varying the ratio of excitatory versus inhibitory neurons, axonal conduction delays and the number of synaptic connections affect a SNN model by observing the change in mean firing rate and polychronization. Our main focus is to examine the effect on the emergence of spatiotemporal time-locked patterns, known as polychronous groups (PNGs). We show that the number of PNGs varies dramatically with a changing proportion of inhibitory neurons, that they increase exponentially as the number of synaptic connections is increased and that they decrease as the maximum axonal delays in the network increases. Our findings show that if we are to use SNNs and PNGs to model cognitive functions we must take into account these critical parameters.

In this paper, a feed forward spiking neural network is tested with spike train patterns with additional and missing spikes. The network is trained with noisy and distorted patterns with an extension of the ReSuMe learning rule to networks with hidden layers. The results show that the multilayer ReSuMe can reliably learn to discriminate highly distorted patterns spanning over 500 ms.

A multiple layers neural network characterized by diverging/converging projections between successive layers activated by an external spatio-temporal input pattern in presence of stochastic background activities was considered. In the previous studies we reported the properties and performance of spike information transmission in the network depending on neuron model type, inputed information type and background activity level. The models were rather simple and can be more detailed and bigger in size for further investigation. Based on a technology developed in the integrated physiology, we have implemented the network model on PhysioDesigner, a platform for multilevel mathematical modeling of physiological systems. This article instructs a use case of PhysioDesigner and the assistive function of PhysioDesigner especially for large size neuronal network modeling is demonstrated.

Most approaches to analysing the spatiotemporal dynamics of neural populations involve binning spike trains. This is likely to underestimate the information carried by spike timing codes, in practice, if they involve high precision patterns of inter-spike intervals (ISIs). In this paper we set out to investigate the differential entropy of multivariate neural spike trains, following the work of Victor. In our framework, the unidimensional special case corresponds to estimating the differential entropy of the ISI distribution; this is generalised to multidimensional cases including patterns across successive ISIs and across cells. We investigated the differential entropy of simulated spike trains with increasing dimensionality, and applied our approach to electrophysiological data recorded from the mouse lateral geniculate nucleus.

The detection and categorization of animate motions is a crucial task underlying social interaction and decision-making. Neural representations of perceived animate objects are built into cortical area STS which is a region of convergent input from intermediate level form and motion representations. Populations of STS cells exist which are selectively responsive to specific action sequences, such as walkers. It is still unclear how and to which extent form and motion information contribute to the generation of such representations and what kind of mechanisms are utilized for the learning processes. The paper develops a cortical model architecture for the unsupervised learning of animated motion sequence representations. We demonstrate how the model automatically selects significant motion patterns as well as meaningful static snapshot categories from continuous video input. Such keyposes correspond to articulated postures which are utilized in probing the trained network to impose implied motion perception from static views. We also show how sequence selective representations are learned in STS by fusing snapshot and motion input and how learned feedback connections enable making predictions about future input. Network simulations demonstrate the computational capacity of the proposed model.

Active exploration is a necessary component of a putative spatial representation system in the mammalian brain. We address the problem of how spatial exploratory behaviour is generated in rodents by combining an artificial neural network model of place coding with a multiobjective evolutionary algorithm that tunes the model parameters so as to maximise the efficiency of environment exploration. A central property of the spatial representation model is an online calibration between external visual cues and path integration, a widely accepted concept in theoretical accounts of spatial learning in animals. We find that the artificially evolved exploration model leads to recurrent patterns of exploratory behaviour in a way observed in experimental studies of spatial exploration in rodents. Our results provide a link between the functional organisation of the biological spatial learning network and the observed high-level patterns of exploratory behaviour.

Neural Oscillators are network of neurons that produce smooth oscillatory/rhythmic pattern. In this paper, a simple network of Matsuoka oscillators is used to generate the bipedal locomotion by coordinating the joint angle using a new method of coordination proposed by Huang. This method makes the control of the joint angle easier as they are related with each other in a sequential manner. Also, this method prevents error accumulation with time. The results show that the robot produced a smooth walking motion.

Following an object’s position relative to oneself is a fundamental functionality required in intelligent real-world interacting robotic systems. This paper presents a computationally efficient vision based 3D tracking system, which can ultimately operate in real-time on autonomous mobile robots in cluttered environments. At the core of the system, two neural inspired event-based dynamic vision sensors (eDVS) independently track a high frequency flickering LED in their respective 2D angular coordinate frame. A self-adjusted feed-forward neural network maps those independent 2D angular coordinates into a Cartesian 3D position in world coordinates. During an initial calibration phase, an object composed of multiple independent markers with known geometry provides relative position information between those markers for network training (without ever using absolute world coordinates for training). In a subsequent application phase tracking a single marker yields position estimates relative to sensor origin, while tracking multiple markers provides additional orientation. The neural inspired vision-based tracking system runs in real-time on ARM7 microcontrollers, without the need for an external PC.

The common approach to estimate the distance of an object in computer vision and robotics is to use stereo vision. Stereopsis, however, provides good estimates only within near space and thus is more suitable for reaching actions. In order to successfully plan and execute an action in far space, other depth cues must be taken into account. Self-body movements, such as head and eye movements or locomotion can provide rich information of depth. This paper proposes a model for integration of static and self-motion-based depth cues for a humanoid robot. Our results show that self-motion-based visual cues improve the accuracy of distance perception and combined with other depth cues provide the robot with a robust distance estimator suitable for both reaching and walking actions.

In this paper, a trajectory tracking control problem for a nonholonomic mobile robot by the integration of a kinematic neural controller (KNC) and a torque neural controller (TNC) is proposed, where both the kinematic and dynamic models contains disturbances. The KNC is a variable structure controller (VSC) based on the sliding mode control theory (SMC), and applied to compensate the kinematic disturbances. The TNC is a inertia-based controller constituted of a dynamic neural controller (DNC) and a robust neural compensator (RNC), and applied to compensate the mobile robot dynamics, and bounded unknown disturbances. Stability analysis with basis on Lyapunov method and simulations results are provided to show the effectiveness of the proposed approach.

The retina performs various key operations on incoming images in order to facilitate higher-level visual processing. Since the retina outperforms existing image enhancing techniques, it follows that computational simulations with biological plausibility are best suited to inform their design and development, as well as help us better understand retina functionality. Recently, it has been determined that quality of vision is dependant on the interaction between rod and cone pathways, traditionally thought to be wholly autonomous. This interaction improves the signal-to-noise ratio (SNR) within the retina and in turn enhances boundary detection by cones. In this paper we therefore propose the first cone simulator that incorporates input from rods. Our results show that rod-cone convergence does improve SNR, therefore allowing for improved contrast sensitivity, and consequently visual perception.

Sand scorpions have a good capability of detecting vibration caused by their prey. They have tactile sense organs in their legs, and they are sensitive to the vibration of surface wave. It is known that the receptor neurons (command neurons) from each leg have inhibitory connections to pinpoint the direction of vibration source, and triad inhibitory connections among receptor neurons have been suggested to explain their orientation behavior. In this paper, we explore the neural network mechanism for the orientation behavior of sand scorpions towards their prey, and by evolving neural networks, we found inhibitory connections among receptor neurons play a significant role for the behavior.

This paper addresses the problem of how dendritic topology and other properties of a neuron can determine its pattern recognition performance. In this study, dendritic trees were evolved using an evolutionary algorithm, which varied both morphologies and other parameters. Based on these trees, we constructed multi-compartmental conductance-based models of neurons. We found that dendritic morphology did have a considerable effect on pattern recognition performance. The results also revealed that the evolutionary algorithm could find effective morphologies, with a performance that was five times better than that of hand-tuned models.

Neural network providing simultaneous perception of features and temporal parameters of sensory stimuli is presented. It is postulated that temporal parameters “when” are processed together with feature parameters “what”, and associations “what–when” are encoded in neocortical areas wherein parameters “what” are perceived. Rate of “internal clock” is inversely proportional to the time of repeated excitation of cortical area via the subthalamic, pedunculopontine and thalamic nuclei. This time depends on dopamine-modulated functioning of parallel cortico–basal ganglia–thalamocortical loop that promote disinhibition of mentioned nuclei and contrasted amplification of firing of cortical neurons initially activated by a stimulus. Parameter “when” is proportional to the number and duration of cycles of repeated cortical excitation. Time counting starts involuntary by a stimulus, or voluntary due to prefrontal cortex activation. Proposed mechanism is identical for visual, auditory or tactile stimuli owing to similarity of functioning of topically organised cortico–basal ganglia–thalamocortical loops.

We investigated the effect of reciprocal connections in a network of modules of simulated spiking neurons. The neural activity is recorded by means of virtual electrodes and EEG-like signals, called electrochipograms (EChG), are analyzed by time- and frequency-domain methods. Bio-inspired processes in the circuits drive the build-up of auto-associative links within each module, which generate an areal activity, recorded by EChG, that reflect the changes in the corresponding functional connectivity within and between neuronal modules. We found that circuits with short inter-layer reciprocal projections exhibited enhanced response as to the stimulus, as to the inner-activity and long inter-layer projections make circuit exhibit non-coherent behavior. We show evidence that all networks of modules are able to process and maintain patterns of activity associated with the stimulus after its offset.

We review the contributions of some known models to the discussion of what the underlying neuronal mechanisms of consciousness creation should be. In particular, we note how different aspects of human mental behavior, such as in psychopathologies and dreams, may contribute to the understanding of these basic components. The interplay of conscious and unconscious mental functioning in the description of the psychoneuroses is analyzed. Aspects such as attentional capabilities, memory functioning, sequentiality and the capacity to create metarepresentations are discussed.

We progressively improve conventional elastic graph matching (EGM) algorithm. In the conventional EGM, each node of a model graph can difficultly detect its corresponding precise position for the most similar Gabor feature extraction on an input low-resolution image. Solving this problem and then finding such a position, we propose a method that the node is allowed to fit among pixels by interpolating aliased Gabor feature representation between the pixels, which is calculated with the others extracted at the neighbor pixels. The model graph can thereby move to the most likely and more precise positions on the input low-resolution image.

We propose a contour operator, called CORF, inspired by the properties of simple cells in visual cortex. It combines, by a weighted geometric mean, the blurred responses of difference-of-Gaussian operators that model cells in the lateral geniculate nucleus (LGN). An operator that has gained particular popularity as a computational model of a simple cell is based on a family of Gabor Functions (GFs). However, the GF operator short-cuts the LGN, and its effectiveness in contour detection tasks, which is assumed to be the primary biological role of simple cells, has never been compared with the effectiveness of alternative operators. We compare the performances of the CORF and the GF operators using the RuG and the Berkeley data sets of natural scenes with associated ground truths. The proposed CORF operator outperforms the GF operator (RuG:

$t(39)\!=\!4.39$

,

$p\!<\! 10^{-4}$

and Berkeley:

$t(499)\!=\!4.95$

,

$p\!<\! 10^{-6}$

).

The success of an ensemble of classifiers depends on the diversity of the underlying features. If a classifier can address more different aspects of the analyzed objects, this allows to improve an ensemble. In this paper we propose an ensemble using as classifier members a Hopfield Neural Network (HNN) that uses Haar-like features as an input template. We analyse the HNN as the only classifier type and also combine it with threshold classifiers to a hybrid neural ensemble, so that the resulting ensemble contains –as members– threshold and neural classifiers. This ensemble architecture is evaluated for the domain of face detection. We show that a HNN that uses summed pixel intensities as input for the classification has the ability to improve the performance of an ensemble.

We analyze the relative relevance of Gabor amplitudes and phases for face recognition. We propose an algorithm to reliably estimate offset point disparities from phase differences and show that disparity-corrected Gabor phase differences are well suited for face recognition in difficult lighting conditions. The method reaches 74.8% recognition rate on the Lighting set of the CAS-PEAL database and 35.7% verification rate on experiment 2.4 of the FRGC database.

This paper describes a probabilistic learning method that is named a contextual probabilistic latent component tree for object and scene categorization. In this method, object classes are obtained by clustering a set of object segments extracted from scene images of each scene category and their categorical co-occurrence relations in scene categories are embedded in the probabilistic latent component tree that is generated as a classification tree of all the object classes of all the scene categories. Through experiments by using images of plural categories in an image database, it is shown that the co-occurrence relation of object categories in scene categories improves performance for object and scene recognition.

A computational model for face feature extraction and recognition capable of achieving a high degree of invariance to illumination and pose is presented. Similar to the complex V1 cells, the model uses a sparse binary code to represent an edge orientation. The binary code represents the face features for recognition. This paper investigates the geometrical structure of the linear space of face representation vectors. For this study the Yale Face Database B is used. It is shown that the biologically inspired procedure provides the face representation of a good quality: vectors representing the faces of the same person under different poses and illumination conditions are grouped together in the vector space. This code enables a very high recognition rate for both the illumination invariance and pose invariance settings.

We recently explored the benefits of a reinforcement learning agent which is supplemented by a symbolic learning level. This second level is represented in the symbolic form of Spohn’s ranking functions. Given this context, we discuss in this paper the creation of symbolic rules from a

Q

-function. We explore several alternatives and show that the rule generation greatly influences the performance of the agent. We provide empirical evidence about which approach to favor. Additionally, the rules created by the considered application are shown to be plausible and understandable.

How does the brain learn to map multi-dimensional sensory inputs to multi-dimensional motor outputs when it can only observe single rewards for the coordinated outputs of the whole network of neurons that make up the brain? We introduce Multi-AGREL, a novel, biologically plausible multi-layer neural network model for multi-dimensional reinforcement learning. We demonstrate that Multi-AGREL can learn non-linear mappings from inputs to multi-dimensional outputs by using only scalar reward feedback. We further show that in Multi-AGREL, the changes in the connection weights follow the gradient that minimizes global prediction error, and that all information required for synaptic plasticity is locally present.

Rhythmic neural circuits play an important role in biological systems in particular in motion generation. They can be entrained by sensory feedback to induce rhythmic motion at a natural frequency, leading to energy-efficient motion. In addition, such circuits can even store the entrained rhythmical patterns through connection weights. Inspired by this, we introduce an adaptive discrete-time neural oscillator system with synaptic plasticity. The system consists of only three neurons and uses adaptive mechanisms based on frequency adaptation and Hebbian-type learning rules. As a result, it autonomously generates periodic patterns and can be entrained by sensory feedback to memorize a pattern. Using numerical simulations we show that this neural system possesses fast and precise convergence behaviour within a wide target frequency range. We use resonant tuning of a pendulum as a simple system for demonstrating possible applications of the adaptive oscillator network.

Extending previous work, we introduce a spiking actor-critic network model of learning from reward and punishment in the basal ganglia. In the model, the striatum is taken to be segregated into populations of medium spiny neurons (MSNs) that carry either D1 or D2 dopamine receptor type. This segregation allows explicit representation of both positive and negative expected outcome within the respective population. In line with recent experiments, we further assume that D1 and D2 MSN populations have opposing dopamine-modulated bidirectional synaptic plasticity. Experiments were conducted in a grid world, where a moving agent had to reach a remote rewarded goal state. The network learned not only to approach the rewarded goal, but also to consequently avoid punishments as opposed to the previous model. The spiking network model explains functional role of D1/D2 MSN segregation within striatum, specifically the reversed direction of dopamine-dependent plasticity found at synapses converging on different MSNs.

We present a Reinforcement Learning (RL)-based model of serotonin which tries to reconcile some of the diverse roles of the neuromodulator. The proposed model uses a novel formulation of utility function, which is a weighted sum of the traditional value function and the risk function. Serotonin is represented by the weightage,

α

, used in this combination. The model is applied to three different experimental paradigms: 1) bee foraging behavior, which involves decision making based on risk, 2) temporal reward prediction task, in which serotonin (

α

) controls the time-scale of reward prediction, and 3) reward/punishment prediction task, in which punishment prediction error depends on serotonin levels. The three diverse roles of serotonin – in time-scale of reward prediction, risk modeling, and punishment prediction – is explained within a single framework by the model.

Neurophysiological experimental results suggest that basal ganglia plays crucial role in action selection while dopamine modifies this process. There are computational models based on these experimental results for action selection. This work focuses on modification of action selection by dopamine release. In the model, a dynamical system is considered for action selection and modification of action selection process is realized by reinforcement learning. The ability of the proposed dynamical system is investigated by bifurcation analysis. Based on the results of this bifurcation analysis, the effect of reinforcement learning on action selection is discussed. The model is implemented on a mobile robot and a foraging task is realized where an exploration in an unfamiliar environment with training in the world is accomplished. Thus, this work fulfills its aim of showing the efficiency of brain-inspired computational models in controlling intelligent agents.

We show how to encode, retrieve and process complex structures equivalent to First-Order Logic (FOL) formulae, with Artificial Neural Networks (ANNs) designed for energy-minimization. The solution constitutes a binding mechanism that uses a neural Working Memory (WM) and a long-term synaptic memory (LTM) that can store both procedural and declarative FOL-like Knowledge-Base (KB). Complex structures stored in LTM are retrieved into the WM only upon need, where they are further processed. The power of our binding mechanism is demonstrated on unification problems: as neurons are dynamically allocated from a pool, most generally unified structures emerge at equilibrium. The network’s size is

O(n

·

k)

, where

n

is the size of the retrieved FOL structures and

k

is the size of the KB. The mechanism is fault-tolerant, as no fatal failures occur when random units fail. The paradigm can be used in a variety of applications, such as language processing, reasoning and planning.

We consider learning a causal ordering of variables in a linear non-Gaussian acyclic model called LiNGAM. Several existing methods have been shown to consistently estimate a causal ordering assuming that all the model assumptions are correct. But, the estimation results could be distorted if some assumptions actually are violated. In this paper, we propose a new algorithm for learning causal orders that is robust against one typical violation of the model assumptions: latent confounders. We demonstrate the effectiveness of our method using artificial data.

In this paper we investigate parameter dependencies in the echo state network (ESN). In particular, we investigate the interplay between reservoir sizes and the choice of the average absolute output feedback connection weight values (

W

OFB

). We consider the multiple sine wave oscillator problem and the powered sine problem. The results show that somewhat contrary to basic intuition (1) smaller reservoir sizes often yield better networks with higher probability; (2) large

W

OFB

values paired with comparatively large reservoirs may strongly decrease the likelihood of generating effective networks; (3) the likelihood of generating an effective ESN depends non-linearly on the choice of

W

OFB

: very small and large weight values often yield higher likelihoods of generating effective ESNs than networks resulting from intermediate

W

OFB

choices. While the considered test problems are rather simple, the insights gained need to be considered when designing effective ESNs for the problem at hand. Nonetheless, further studies appear necessary to be able to explain the actual reasons behind the observed phenomena.

When an echo state network with feedback connections is trained with teacher forcing and later run in free mode, one often gets problems with stability. In this paper an echo state network is trained to execute an arm movement. A sequence with the desired coordinates of the limbs in each time step is provided to the network together with the current limb coordinates. The network must find the appropriate angle velocities that will keep the arms on this trajectory. The current limb coordinates are indirect feedback from the motor output via the simulator. We do get a problem with stability in this setup. One simple remedy is adding noise to the internal states of the network. We verify that this helps, but we also suggest a new training strategy that leeds to even better performance on this task.

In this paper, we propose the Ensemble-Recursive-Rule eXtraction (E-Re-RX) algorithm, which is a rule extraction method from ensemble neural networks. We demonstrate that the use of ensemble neural networks produces higher recognition accuracy than individual neural networks and the extracted rules are more comprehensible. E-Re-RX algorithm is an effective rule extraction algorithm for dealing with data sets that mix discrete and continuous attributes. In this algorithm, primary rules are generated as well as secondary rules to handleonlythoseinstances that do not satisfy the primary rules, and then these rules are integrated. We show that this reduces the complexity of using multiple neural networks. This method achieves extremely high recognition rates, even with multiclass problems.

The accuracy with which the brain can infer the value of a stimulus variable depends on both the amount of stimulus information that is represented in sensory neurons (encoding) and the mechanism by which this information is subsequently retrieved from the responses of these neurons (decoding). Previous studies have mainly focused on either the encoding or the decoding aspect. Here, we present a new framework that functionally links the two. More specifically, we demonstrate that optimal (efficient) population codes which guarantee uniform firing rate distributions allow the accurate emulation of optimal (Bayesian) inference using a biophysically plausible neural mechanism. The framework provides predictions for estimation bias and variability as a function of stimulus prior, strength and integration time, as well as physiological parameters such as tuning curves and spontaneous firing rates. Our framework represents an example of the duality between representation and computation in neural information processing.

Recurrent neural networks (RNNs) in combination with a pooling operator and the neighbourhood components analysis (NCA) objective function are able to detect the characterizing dynamics of sequences and embed them into a fixed-length vector space of arbitrary dimensionality. Subsequently, the resulting features are meaningful and can be used for visualization or nearest neighbour classification in linear time. This kind of metric learning for sequential data enables the use of algorithms tailored towards fixed length vector spaces such as ℝ

n

.

The visual system processes the features and movement of an object in separate pathways, called the ventral and dorsal streams. To integrate this principle in a functional model, a recurrent predictive network with a horizontal product is introduced. Learned in an unsupervised manner, two sets of hidden units represent cells in the ventral and dorsal pathways, respectively. Experiments show that the activity in the ventral-like units persists, given that the same feature appears in the receptive field, whilst the activity in the dorsal-like units shows a fluctuating pattern with different directions of object movements. Moreover, we show that the position information predicts the input’s future position taking into account its moving direction due to the direction-selective responses of the dorsal-like units.

Generating stable yet performant spiking neural reservoirs for classification applications is still an open issue. This is due to the extremely non-linear dynamics of recurrent spiking neural networks. In this perspective, a local and unsupervised learning rule that tunes the reservoir toward self-organized criticality is proposed, and applied to networks of leaky integrate-and-fire neurons with random and small-world topologies. Longer sustained activity for both topologies was elicited after learning compared to spectral radius normalization (global rescaling scheme). The ability to control the desired regime of the reservoir was shown and quick convergence toward it was observed for speech signals. Proposed regulation method can be applied online and leads to reservoirs more strongly adapted to the task at hand.

Recent research has revealed that hierarchical linguistic structures can emerge in a recurrent neural network with a sufficient number of delayed context layers. As a representative of this type of network the Multiple Timescale Recurrent Neural Network (MTRNN) has been proposed for recognising and generating known as well as unknown linguistic utterances. However the training of utterances performed in other approaches demands a high training effort. In this paper we propose a robust mechanism for adaptive learning rates and internal states to speed up the training process substantially. In addition we compare the generalisation of the network for the adaptive mechanism as well as the standard fixed learning rates finding at least equal capabilities.

In modern society, the more complex information and technology become, the more important data analysis become. In particular, a data, which has a variety of elements, is complex, and it is extremely difficult to estimate the state which generates data from observed data. To handle those hidden states, we propose an appropriate model using Spherical-Self Organizing Map (S-SOM) with Hidden Markov Model (HMM) which can estimate the hidden state.

In the present work we showed that together with improved stability the Intrinsic Plasticity (IP) tuned Echo State Network (ESN) reservoirs possess also better clustering abilities that opens a possibility for application of ESNs in multidimensional data clustering. The revealed ability of ESNs is demonstrated first on an artificially created data set with known in advance number and position of clusters. Automated procedure for multidimensional data clustering was proposed. It allows discovering multidimensional data structure without specification in advance the clusters number. The developed procedure was further applied to a real data set containing concentrations of three alloying elements in numerous steel compositions. The obtained number and position of clusters showed logical from the practical point of view data separation.

Motion perception is theoretically understood as the detection of sequential optical changes at two locations in the visual array. Experiments on generalized apparent motion have demonstrated, however, that sequentiality is not necessarily required for the detection of motion [1] leading to instantaneous counter-change as an alternative theoretical view of motion detection [2]. Here we generalize the counter-change model to spatially and temporally continuous motion. Transients detected within the receptive fields of edge filters are combined in a space-time continuous neural dynamics, in which nonlinear neural interaction leads to the detection decision. We show that the model enables the detection of continuous motion while also accounting for psychophysical results on generalized apparent motion.

We investigate how unsupervised training of recurrent neural networks (RNNs) and their deep hierarchies can benefit a supervised task like temporal pattern detection. The RNNs are fully and fast trained by unsupervised algorithms and only supervised feed-forward readouts are used. The unsupervised RNNs are shown to perform better in a rigorous comparison against state-of-art random reservoir networks. Unsupervised greedy bottom-up trained hierarchies of such RNNs are shown being capable of big performance improvements over single layer setups.

Previous words in the sentence can influence the processing of the current word in the timescale of hundreds of milliseconds. The current research provides a possible explanation of how certain aspects of this on-line language processing can occur, based on the dynamics of recurrent cortical networks. We simulate prefrontal area BA47 as a recurrent network that receives on-line input of “grammatical” words during sentence processing, with plastic connections between cortex and striatum (homology with Reservoir Computing). The system is trained on sentence-meaning pairs, where meaning is coded as activation in the striatum corresponding to the roles that different “semantic words” play in the sentences. The model learns an extended set of grammatical constructions, and demonstrates the ability to generalize to novel constructions. This demonstrates that a RNN can decode grammatical structure from sentences in an on-line manner in order to generate a predictive representation of the meaning of the sentences.

In this paper, we propose a mechanism to effectively control the overall neural activity in the reservoir of a Liquid State Machine (LSM) in order to achieve both a high sensitivity of the reservoir to weak stimuli as well as an improved resistance to over-stimulation for strong inputs. The idea is to employ a mechanism that dynamically changes the firing threshold of a neuron in dependence of its spike activity. We experimentally demonstrate that reservoirs employing this neural model significantly increase their separation capabilities. We also investigate the role of dynamic and static synapses in this context. The obtained results may be very valuable for LSM based real-world application in which the input signal is often highly variable causing problems of either too little or too much network activity.

In this paper we define a kernel function which is the dual space equivalent of infinitely large sparse threshold unit networks. We first explain how to couple a kernel function to an infinite recurrent neural network, and next we use this definition to apply the theory to sparse threshold unit networks. We validate this kernel function with a theoretical analysis and an illustrative signal processing task.

Unsupervised learning of feature hierarchies is often a good initialization for supervised training of deep architectures. In existing deep learning methods, these feature hierarchies are built layer by layer in a greedy fashion using auto-encoders or restricted Boltzmann machines. Both yield encoders, which compute linear projections followed by a smooth thresholding function. In this work, we demonstrate that these encoders fail to find stable features when the required computation is in the exclusive-or class. To overcome this limitation, we propose a two-layer encoder which is not restricted in the type of features it can learn. The proposed encoder can be regularized by an extension of previous work on contractive regularization. We demonstrate the advantages of two-layer encoders qualitatively, as well as on commonly used benchmark datasets.

A common technique in visual object recognition is to use a sparse encoding of low-level input with a feature dictionary followed by a spatial pooling over local neighbourhoods. While some methods stack these in alternating layers within hierarchies, using these two stages alone can also produce state-of-the-art results. Following from vision, this framework is moving in to speech and audio processing tasks. We investigate the effect of architectural choices when applied to a spoken digit recognition task. We find that the unsupervised learning of features has a negligible effect on the classification, with the number of and size of the features being a greater determinant for recognition. Finally, we show that, given an optimised architecture, sparse coding performs comparably with Hidden Markov Models (HMMs) and outperforms K-means clustering.

We interpret biological motion trajectories as composed of sequences of sub-blocks or

motion primitives

. Such primitives, together with the information,

when

they occur during a motion, provide a compact representation of movement. We present a two-layer model for movement generation, where the higher level consists of a number of spiking neurons that trigger motion primitives in the lower level. Given a set of handwritten character trajectories, we learn motion primitives, together with the timing information, with a variant of shift-NMF that is able to cope with large data sets. From the timing information for a class of characters, we then learn a generative model based on a stochastic Integrate-and-Fire neuron model. We show that we can generate good reconstructions of characters with shared primitives for all characters modeled.

In this work, we introduce the kinetic framework for modelling synaptic transmission in an existing neural mass model of the thalamocortical circuitry to study Electroencephalogram (EEG) slowing within the alpha frequency band (8–13 Hz), a hallmark of Alzheimer’s disease (AD). Ligand-gated excitatory and inhibitory synapses mediated by AMPA (

α

-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid) and GABA

A

(gamma-amino-butyric acid) receptors respectively are modelled. Our results show that the concentration of the GABA neurotransmitter acts as a bifurcation parameter, causing the model to switch from a limit cycle mode to a steady state. Further, the retino-geniculate pathway connectivity plays a significant role in modulating the power within the alpha band, thus conforming to research proposing ocular biomarkers in AD. Overall, kinetic modelling of synaptic transmission in neural mass models has enabled a more detailed investigation into the neural correlates underlying abnormal EEG in AD.

In the last 20 years a lot of works in literature analysed and proposed several methods capable to predict the occurrence of seizures from the electroencephalogram (EEG) of epileptic patients. One of the best results was obtained using a version of the maximum Lyapunov exponent (Lmax) for predicting the advent of a seizure, but in spite of promising results presented, more recent evaluations could not reproduce these optimistic findings. Following this trend, in this paper we propose a new integrative technique starting from two different paradigms: Chaos and Neural Networks (NN). The new framework has been tested on long term intracerebral stereo-EEG (sEEG) recordings, with very good results. We present this way of analysis as the key for modelling brain mechanisms during epileptic seizures, going over critical state reported in literature for Lmax with suited computational methods, providing theoretical justifications. This is a preliminary work with numerous possible evolutions.

Brain Computer Interface spellers based on the P300 paradigm traditionally use a fixed number of epochs (stimulus presentations) to predict a letter. In this contribution, we introduce a dynamical adjustment of the number of epochs based on a threshold on the confidence of a probabilistic classifier. This allows the average required number of epochs to be lowered drastically. As such, using a conceptually simple modification with no impact on computational requirements, we obtain a P300 speller which is not only faster but also more accurate, which in turn increases the usability of the system substantially.

One problem in current Brain-Computer Interfaces (BCIs) is non-stationarity of the underlying signals. This causes deteriorating performance throughout a session and difficulties to transfer a classifier from one session to another, which results in the need of collecting training data every session. Using an adaptive classifier is one solution to keep the performance stable and reduce the amount of training that is needed for a good BCI performance. In this paper we present an approach for an adaptive classifier based on a Support Vector Machine (SVM). We evaluate its advantage on offline BCI data and show its benefits and online feasibility in an online experiment using a MEG-based BCI with 10 subjects.

In the field of human computer interaction (HCI) the detection and classification of human activity patterns has become an important challenge. The problem can be understood as a specific problem of pattern recognition which addresses three topics, namely fusion of multiple modalities, spatio-temporal structures and a vast variety of pattern appearances the more abstract a pattern gets. In order to approach the problem, we propose a layered architecture which decomposes temporal patterns into elementary sub-patterns. Within each layer the patterns are detected using Markov models. The results of a layer are passed to the next successive layer such that on each layer the temporal granularity and the complexity of patterns increases. A dataset containing activities in an office scenario was recorded. The activities are decomposed to basic actions which are detected on the first layer. We evaluated a two-layered architecture using the dataset showing the feasibility of the approach.

Reservoir Computing is a paradigm of artificial neural networks that has obtained promising results. However there are some disadvantages: the reservoir is created randomly and needs to be large enough to be able to capture all the features of the data. For this work we use PSO – Particle Swarm Optimization to optimize the initial parameters of the Reservoir Computing. The results obtained with the optimization method are compared with results obtained by an exhaustive search for global parameters generation of Reservoir Computing. Five time series were used to show that the optimization method reduces the number of training cycles required to train the system.

In this paper we propose a novel method for one-class classification. The proposed method analyses the limit of all feature dimensions to find the true border which describes the normal class. To this end, it simulates the novelty class by creating artificial prototypes outside the normal description. The parameters involved in the definition of the border are optimized via particle swarm optimization (PSO), which enables the method to describe data distributions with complex shapes. An experimental analysis is conducted with the proposed method using twelve data sets and considering the performance measures (i) Area Under the ROC Curve (AUC), (ii) training time, and (iii) prototype reduction. A comparison with One-Class SVM (OCSVM), kMeansDD, ParzenDD and kNNDD is carried out. The results show that performance of the proposed method is equivalent to OCSVM regarding the AUC, yet the proposed method outperforms OCSVM regarding the number of stored prototypes and training time.

This paper presents the use of a bi-objective genetic algorithm to select attributes for an ensemble system. This is achieved by using this technique to simultaneously maximize the individual diversity of the base classifiers and the group diversity of an ensemble system. In order to evaluate the possible solutions obtained by this technique, two filter-based evaluation criteria will be used. Filter-based criteria were chosen because they are independent of the learning algorithm and have a low computational cost.

As machine learning acquires special attention for real-world problem solving, a growing number of new problems not previously considered have appeared. One of such problems is the imbalance in class distributions, which is said to hinder the performance of traditional error-minimization-based classification algorithms. In this paper we propose an improved rule-based decision boundary for the Support Vector Domain Description that uses an additional nested classification unit to improve the accuracy of the outlier class, hence improving the overall performance of the classifier. Computer simulations show that the proposed strategy, which we have termed Dual Support Vector Domain Description, outperforms related literature approaches in several benchmark instances.

Although the individuals’ transport behavioural modelling is a complex task, it has a notable social and economic impact. Thus, in this paper Fuzzy Cognitive Maps are explored to represent the behaviour and operation of such systems. This technique allows modelling how the travellers make decisions based on their knowledge of different transport modes properties at different levels of abstraction. We use learning of Fuzzy Cognitive Maps to describe travellers’ behaviour and change trends in different abstraction levels. The results of this study will help transportation policy decision makers in better understanding of people’s needs and consequently will help them actualizing different policy formulations and implementations.

We describe a computational model of motor areas of the cerebral cortex. The model combines Bayesian networks, competitive learning and reinforcement learning. We found that decision-making using MPE (Most Probable Explanation) approximates the ideal decision-making in this model, which suggests that MPE calculation is a promising model of not only sensory-cortex recognition, already addressed by previous works, but also motor-cortex decision-making.