Reconeixement de veu mitjançant xarxes neuronals

La interacció home-màquina per mitjà de la veu cobreix moltes àrees d’investigació. Es destaquen entre altres, el reconeixement de la parla, la síntesis i identificació de discurs, la verificació i identificació de locutor i l’activació per veu (ordres) de sistemes robòtics. Reconèixer la parla és natural i simple per a les persones, però és un treball complex per a les màquines, pel qual existeixen diverses metodologies i tècniques, entre elles les Xarxes Neuronals. L’objectiu d’aquest treball és desenvolupar una eina en Matlab per al reconeixement i identificació de paraules pronunciades per un locutor, entre un conjunt de paraules possibles, i amb una bona fiabilitat dins d’uns marges preestablerts. El sistema és independent del locutor que pronuncia la paraula, és a dir, aquest locutor no haurà intervingut en el procés d’entrenament del sistema. S’ha dissenyat una interfície que permet l’adquisició del senyal de veu i el seu processament mitjançant xarxes neuronals i altres tècniques. Adaptant una part de control al sistema, es podria utilitzar per donar ordres a un robot com l’Alfa6Uvic o qualsevol altre dispositiu.

Reconeixement de veu mitjançant xarxes neuronals

La interacció home-màquina per mitjà de la veu cobreix moltes àrees d’investigació. Es destaquen entre altres, el reconeixement de la parla, la síntesis i identificació de discurs, la verificació i identificació de locutor i l’activació per veu (ordres) de sistemes robòtics. Reconèixer la parla és natural i simple per a les persones, però és un treball complex per a les màquines, pel qual existeixen diverses metodologies i tècniques, entre elles les Xarxes Neuronals. L’objectiu d’aquest treball és desenvolupar una eina en Matlab per al reconeixement i identificació de paraules pronunciades per un locutor, entre un conjunt de paraules possibles, i amb una bona fiabilitat dins d’uns marges preestablerts. El sistema és independent del locutor que pronuncia la paraula, és a dir, aquest locutor no haurà intervingut en el procés d’entrenament del sistema. S’ha dissenyat una interfície que permet l’adquisició del senyal de veu i el seu processament mitjançant xarxes neuronals i altres tècniques. Adaptant una part de control al sistema, es podria utilitzar per donar ordres a un robot com l’Alfa6Uvic o qualsevol altre dispositiu.

Scene classifier for real time road detection

L’objectiu principal del projecte és el de classificar escenes de carretera en funció del contingut de les imatges per així poder fer un desglossament sobre quin tipus de situació tenim en el moment. És important que fixem els paràmetres necessaris en funció de l’escenari en què ens trobem per tal de treure el màxim rendiment possible a cada un dels algoritmes. La seva funcionalitat doncs, ha de ser la d’avís i suport davant els diferents escenaris de conducció. És a dir, el resultat final ha de contenir un algoritme o aplicació capaç de classificar les imatges d’entrada en diferents tipus amb la màxima eficiència espacial i temporal possible. L’algoritme haurà de classificar les imatges en diferents escenaris. Els algoritmes hauran de ser parametritzables i fàcilment manejables per l’usuari. L’eina utilitzada per aconseguir aquests objectius serà el MATLAB amb les toolboxs de visió i xarxes neuronals instal·lades.

A numerical solver for a nonlinear Fokker-Planck equation representation of neuronal network dynamics

To describe the collective behavior of large ensembles of neurons in neuronal network, a kinetic theory description was developed in [13, 12], where a macroscopic representation of the network dynamics was directly derived from the microscopic dynamics of individual neurons, which are modeled by conductance-based, linear, integrate-and-fire point neurons. A diffusion approximation then led to a nonlinear Fokker-Planck equation for the probability density function of neuronal membrane potentials and synaptic conductances. In this work, we propose a deterministic numerical scheme for a Fokker-Planck model of an excitatory-only network. Our numerical solver allows us to obtain the time evolution of probability distribution functions, and thus, the evolution of all possible macroscopic quantities that are given by suitable moments of the probability density function. We show that this deterministic scheme is capable of capturing the bistability of stationary states observed in Monte Carlo simulations. Moreover, the transient behavior of the firing rates computed from the Fokker-Planck equation is analyzed in this bistable situation, where a bifurcation scenario, of asynchronous convergence towards stationary states, periodic synchronous solutions or damped oscillatory convergence towards stationary states, can be uncovered by increasing the strength of the excitatory coupling. Finally, the computation of moments of the probability distribution allows us to validate the applicability of a moment closure assumption used in [13] to further simplify the kinetic theory.

Efficient learning of reactive robot behaviors with a Neural-Q_learning approach

The purpose of this paper is to propose a Neural-Q_learning approach designed for online learning of simple and reactive robot behaviors. In this approach, the Q_function is generalized by a multi-layer neural network allowing the use of continuous states and actions. The algorithm uses a database of the most recent learning samples to accelerate and guarantee the convergence. Each Neural-Q_learning function represents an independent, reactive and adaptive behavior which maps sensorial states to robot control actions. A group of these behaviors constitutes a reactive control scheme designed to fulfill simple missions. The paper centers on the description of the Neural-Q_learning based behaviors showing their performance with an underwater robot in a target following task. Real experiments demonstrate the convergence and stability of the learning system, pointing out its suitability for online robot learning. Advantages and limitations are discussed

A behavior-based scheme using reinforcement learning for autonomous underwater vehicles

This paper presents a hybrid behavior-based scheme using reinforcement learning for high-level control of autonomous underwater vehicles (AUVs). Two main features of the presented approach are hybrid behavior coordination and semi on-line neural-Q_learning (SONQL). Hybrid behavior coordination takes advantages of robustness and modularity in the competitive approach as well as efficient trajectories in the cooperative approach. SONQL, a new continuous approach of the Q_learning algorithm with a multilayer neural network is used to learn behavior state/action mapping online. Experimental results show the feasibility of the presented approach for AUVs

Avenços en l'estudi de la litiasi renal: mètodes de chemical imaging i intel·ligencia artificial (ANN) per la caracterització de càlculs renals

La litiasi urinària és un trastorn que implica la formació de precipitats en qualsevol part del tracte urinari. Aquest és un desordre comú que afecta aproximadament a una desena part de la població de la Unió Europea al llarg de la seva vida. A més, durant els cinc anys posteriors a un episodi litiàsic el percentatge de recurrència dels pacients és del 45 al 75%. Aquest trastorn urinari està influït per una gran quantitat de variables, d’origen fisiològic, psicològic i ambiental. Els episodis litiàsics, es poden solucionar espontàniament, amb l’expulsió del càlcul renal, o bé a través de diverses intervencions mèdiques. Els tractaments mèdics derivats de la litiasi urinària; és a dir, la fragmentació del càlcul, intervencions quirúrgiques i tractaments posteriors generen unes grans despeses als sistemes mèdics. Pels motius exposats, la identificació del desordre que ha originat l’episodi litiàsic és de radical importància, per tal de minimitzar el risc de reincidència. Els mètodes més usuals per determinar les causes que desencadenen la formació del càlcul renal són les anàlisis d’orina i l’estudi del càlcul generat. La correcta descripció de la composició i, especialment, l’estructura del càlcul renal pot aportar informació clau sobre les possibles causes de la seva formació, tant de l’inici de nucleació del càlcul com de les successives etapes de creixement cristal·lí. Tenint en compte aquest darrer aspecte, el present estudi s’ha dirigit a la caracterització de càlculs urinaris mitjançant l’aplicació de metodologies d’imatge química (Hyperspectral Imaging), el que va contribuir a determinar les principals característiques espectrals de cada compost majoritari als càlculs renals. D’altra banda, la utilització de mostres de composició coneguda ha possibilitat la creació d’un model amb Xarxes Neuronals Artificials, que permet la classificació de noves mostres de composició desconeguda, de manera més ràpida que el procediment actual. Aquest treball constitueix una nova contribució a la comprensió de l’estructura de les pedres de ronyó, així com les condicions de la seva formació. Els resultats obtinguts destaquen les possibilitats que presenten les tècniques emprades al camp de la litiasi renal, que permeten complementar els coneixements existents enfocats a millorar la qualitat de vida dels pacients.

Volatility smile extrapolation with an artificial neural network

I use a multi-layer feedforward perceptron, with backpropagation learning implemented via stochastic gradient descent, to extrapolate the volatility smile of Euribor derivatives over low-strikes by training the network on parametric prices.

Impairment of mossy fiber long-term potentiation and associative learning in pituitary adenylate cyclase activating polypeptide type I receptor-deficient Mice

The pituitary adenylate cyclase activating polypeptide (PACAP) type I receptor (PAC1) is a G-protein-coupled receptor binding the strongly conserved neuropeptide PACAP with 1000-fold higher affinity than the related peptide vasoactive intestinal peptide. PAC1-mediated signaling has been implicated in neuronal differentiation and synaptic plasticity. To gain further insight into the biological significance of PAC1-mediated signaling in vivo, we generated two different mutant mouse strains, harboring either a complete or a forebrain-specific inactivation of PAC1. Mutants from both strains show a deficit in contextual fear conditioning, a hippocampus-dependent associative learning paradigm. In sharp contrast, amygdala-dependent cued fear conditioning remains intact. Interestingly, no deficits in other hippocampus-dependent tasks modeling declarative learning such as the Morris water maze or the social transmission of food preference are observed. At the cellular level, the deficit in hippocampus-dependent associative learning is accompanied by an impairment of mossy fiber long-term potentiation (LTP). Because the hippocampal expression of PAC1 is restricted to mossy fiber terminals, we conclude that presynaptic PAC1-mediated signaling at the mossy fiber synapse is involved in both LTP and hippocampus-dependent associative learning.

Bradykinin, but not muscarinic, inhibition of M-current in rat sympathetic ganglion neurons involves phospholipase C-β4

Rat superior cervical ganglion (SCG) neurons express low-threshold noninactivating M-type potassium channels (I-K(M)), which can be inhibited by activation of M-1 muscarinic receptors (M-1 mAChR) and bradykinin (BK) B-2 receptors. Inhibition by the M1 mAChR agonist oxotremorine methiodide (Oxo-M) is mediated, at least in part, by the pertussis toxin-insensitive G-protein G alpha (q) (Caulfield et al., 1994; Haley et al., 1998a), whereas BK inhibition involves G alpha (q) and/or G alpha (11) (Jones et al., 1995). G alpha (q) and G alpha (11) can stimulate phospholipase C-beta (PLC-beta), raising the possibility that PLC is involved in I-K(M) inhibition by Oxo-M and BK. RT-PCR and antibody staining confirmed the presence of PLC-beta1, - beta2, - beta3, and - beta4 in rat SCG. We have tested the role of two PLC isoforms (PLC-beta1 and PLC-beta4) using antisense-expression constructs. Antisense constructs, consisting of the cytomegalovirus promoter driving antisense cRNA corresponding to the 3'-untranslated regions of PLC-beta1 and PLC-beta4, were injected into the nucleus of dissociated SCG neurons. Injected cells showed reduced antibody staining for the relevant PLC-beta isoform when compared to uninjected cells 48 hr later. BK inhibition of I-K(M) was significantly reduced 48 hr after injection of the PLC-beta4, but not the PLC-beta1, antisense-encoding plasmid. Neither PLC-beta antisense altered M-1 mAChR inhibition by Oxo-M. These data support the conclusion of Cruzblanca et al. (1998) that BK, but not M-1 mAChR, inhibition of I-K(M) involves PLC and extends this finding by indicating that PLC-beta4 is involved.

Impaired fast-spiking, suppressed cortical inhibition and increased susceptibility to seizures in mice lacking Kv3.2 K+ channel proteins

Voltage-gated K+ channels of the Kv3 subfamily have unusual electrophysiological properties, including activation at very depolarized voltages (positive to −10 mV) and very fast deactivation rates, suggesting special roles in neuronal excitability. In the brain, Kv3 channels are prominently expressed in select neuronal populations, which include fast-spiking (FS) GABAergic interneurons of the neocortex, hippocampus, and caudate, as well as other high-frequency firing neurons. Although evidence points to a key role in high-frequency firing, a definitive understanding of the function of these channels has been hampered by a lack of selective pharmacological tools. We therefore generated mouse lines in which one of the Kv3 genes, Kv3.2, was disrupted by gene-targeting methods. Whole-cell electrophysiological recording showed that the ability to fire spikes at high frequencies was impaired in immunocytochemically identified FS interneurons of deep cortical layers (5-6) in which Kv3.2 proteins are normally prominent. No such impairment was found for FS neurons of superficial layers (2-4) in which Kv3.2 proteins are normally only weakly expressed. These data directly support the hypothesis that Kv3 channels are necessary for high-frequency firing. Moreover, we found that Kv3.2 −/− mice showed specific alterations in their cortical EEG patterns and an increased susceptibility to epileptic seizures consistent with an impairment of cortical inhibitory mechanisms. This implies that, rather than producing hyperexcitability of the inhibitory interneurons, Kv3.2 channel elimination suppresses their activity. These data suggest that normal cortical operations depend on the ability of inhibitory interneurons to generate high-frequency firing.

K(+) channel expression distinguishes subpopulations of parvalbumin- and somatostatin-containing neocortical interneurons

Kv3.1 and Kv3.2 K+ channel proteins form similar voltage-gated K+ channels with unusual properties, including fast activation at voltages positive to −10 mV and very fast deactivation rates. These properties are thought to facilitate sustained high-frequency firing. Kv3.1 subunits are specifically found in fast-spiking, parvalbumin (PV)-containing cortical interneurons, and recent studies have provided support for a crucial role in the generation of the fast-spiking phenotype. Kv3.2 mRNAs are also found in a small subset of neocortical neurons, although the distribution of these neurons is different. We raised antibodies directed against Kv3.2 proteins and used dual-labeling methods to identify the neocortical neurons expressing Kv3.2 proteins and to determine their subcellular localization. Kv3.2 proteins are prominently expressed in patches in somatic and proximal dendritic membrane as well as in axons and presynaptic terminals of GABAergic interneurons. Kv3.2 subunits are found in all PV-containing neurons in deep cortical layers where they probably form heteromultimeric channels with Kv3.1 subunits. In contrast, in superficial layer PV-positive neurons Kv3.2 immunoreactivity is low, but Kv3.1 is still prominently expressed. Because Kv3.1 and Kv3.2 channels are differentially modulated by protein kinases, these results raise the possibility that the fast-spiking properties of superficial- and deep-layer PV neurons are differentially regulated by neuromodulators. Interestingly, Kv3.2 but not Kv3.1 proteins are also prominent in a subset of seemingly non-fast-spiking, somatostatin- and calbindin-containing interneurons, suggesting that the Kv3.1–Kv3.2 current type can have functions other than facilitating high-frequency firing.

From SNPs to pathways: integration of functional effect of sequence variations on models of cell signalling pathways

Background: Single nucleotide polymorphisms (SNPs) are the most frequent type of sequence variation between individuals, and represent a promising tool for finding genetic determinants of complex diseases and understanding the differences in drug response. In this regard, it is of particular interest to study the effect of non-synonymous SNPs in the context of biological networks such as cell signalling pathways. UniProt provides curated information about the functional and phenotypic effects of sequence variation, including SNPs, as well as on mutations of protein sequences. However, no strategy has been developed to integrate this information with biological networks, with the ultimate goal of studying the impact of the functional effect of SNPs in the structure and dynamics of biological networks. Results: First, we identified the different challenges posed by the integration of the phenotypic effect of sequence variants and mutations with biological networks. Second, we developed a strategy for the combination of data extracted from public resources, such as UniProt, NCBI dbSNP, Reactome and BioModels. We generated attribute files containing phenotypic and genotypic annotations to the nodes of biological networks, which can be imported into network visualization tools such as Cytoscape. These resources allow the mapping and visualization of mutations and natural variations of human proteins and their phenotypic effect on biological networks (e.g. signalling pathways, protein-protein interaction networks, dynamic models). Finally, an example on the use of the sequence variation data in the dynamics of a network model is presented. Conclusion: In this paper we present a general strategy for the integration of pathway and sequence variation data for visualization, analysis and modelling purposes, including the study of the functional impact of protein sequence variations on the dynamics of signalling pathways. This is of particular interest when the SNP or mutation is known to be associated to disease. We expect that this approach will help in the study of the functional impact of disease-associated SNPs on the behaviour of cell signalling pathways, which ultimately will lead to a better understanding of the mechanisms underlying complex diseases.

What can spike train distances tell us about the neural code?

Time scale parametric spike train distances like the Victor and the van Rossum distancesare often applied to study the neural code based on neural stimuli discrimination.Different neural coding hypotheses, such as rate or coincidence coding,can be assessed by combining a time scale parametric spike train distance with aclassifier in order to obtain the optimal discrimination performance. The time scalefor which the responses to different stimuli are distinguished best is assumed to bethe discriminative precision of the neural code. The relevance of temporal codingis evaluated by comparing the optimal discrimination performance with the oneachieved when assuming a rate code.We here characterize the measures quantifying the discrimination performance,the discriminative precision, and the relevance of temporal coding. Furthermore,we evaluate the information these quantities provide about the neural code. Weshow that the discriminative precision is too unspecific to be interpreted in termsof the time scales relevant for encoding. Accordingly, the time scale parametricnature of the distances is mainly an advantage because it allows maximizing thediscrimination performance across a whole set of measures with different sensitivitiesdetermined by the time scale parameter, but not due to the possibility toexamine the temporal properties of the neural code.

Neuronal adaptation effects in decision making

Recently, there has been an increased interest on the neural mechanisms underlying perceptual decision making. However, the effect of neuronal adaptation in this context has not yet been studied. We begin our study by investigating how adaptation can bias perceptual decisions. We considered behavioral data from an experiment on high-level adaptation-related aftereffects in a perceptual decision task with ambiguous stimuli on humans. To understand the driving force behind the perceptual decision process, a biologically inspired cortical network model was used. Two theoretical scenarios arose for explaining the perceptual switch from the category of the adaptor stimulus to the opposite, nonadapted one. One is noise-driven transition due to the probabilistic spike times of neurons and the other is adaptation-driven transition due to afterhyperpolarization currents. With increasing levels of neural adaptation, the system shifts from a noise-driven to an adaptation-driven modus. The behavioral results show that the underlying model is not just a bistable model, as usual in the decision-making modeling literature, but that neuronal adaptation is high and therefore the working point of the model is in the oscillatory regime. Using the same model parameters, we studied the effect of neural adaptation in a perceptual decision-making task where the same ambiguous stimulus was presented with and without a preceding adaptor stimulus. We find that for different levels of sensory evidence favoring one of the two interpretations of the ambiguous stimulus, higher levels of neural adaptation lead to quicker decisions contributing to a speed–accuracy trade off.