Hebbian Learning Mechanisms Help Explain the Maturation of Multisensory Speech Integration in Children with Autism Spectrum Disorder (ASD) and with Typical Development (TD): a Neurocomputational Analysis. Cristiano Cuppini (cristiano.cuppini@unibo.it) Mauro Ursino (mauro.ursino@unibo.it) Elisa Magosso (elisa.magosso@unibo.it) Department of Electric, Electronic and Information Engineering, University of Bologna, 2 Viale Risorgimento Bologna, 40136, Italy Lars A. Ross (lars.ross@einstein.yu.edu) John J. Foxe (john.foxe@einstein.yu.edu) Sophie Molholm (sophie.molholm@einstein.yu.edu) Department of Pediatrics and Neuroscience, Albert Einstein College of Medicine, 1225 Morris Park Avenue Bronx, NY 10461, USA Abstract significantly improved when one can see the speaker’s articulations. Accordingly, the appropriate development of Cognitive tasks such as communication and speech comprehension rely on the brain’s ability to exploit and multisensory speech integration (MSI) greatly affects a integrate sensory information of different modalities. child’s ability to relate with others. Ample experimental Accordingly, the appropriate development of multisensory evidence has shown that MSI appears to be highly immature speech integration (MSI) greatly influences a child’s ability to at birth and that continues to develop late into childhood successfully relate with others. Several experimental findings (Brandwein et al., 2010). Moreover, children with autism have shown that speech intelligibility is affected by spectrum disorder (ASD) presenting impaired MSI early in visualizing a speaker’s articulations, and that MSI continues developing late into childhood. This work aims at developing life, show an amelioration in the adolescent years (de Boer- a network to analyze the role of the sensory experience during Schellekens et al., 2013; Foxe et al., 2015). These evidences the early stages of life, as a mechanism responsible for the suggest that there may be delays in the maturation of MSI maturation of these integrative abilities in teenagers. We for children with ASD that resolve at this point. Multiple extended a model realized to study multisensory integration in studies have shown multisensory processing deficits in ASD cortical regions (Magosso et al., 2012; Cuppini et al, 2014) by in the absence of comparable unisensory deficits, suggesting incorporating a multisensory area known to be involved in that they represent impairment of neural processes that have audiovisual speech processing, the superior temporal sulcus (STS). The model suggests that the maturation of MSI is direct and specific impact on MSI. However, the neural primarily due to the maturation of direct connections among basis of the impairment remains unknown. primary unisensory regions. This process was the results of a A region of particular interest for the maturation of MSI is training phase during which the network was exposed to the superior temporal sulcus (STS), an association cortex sensory-specific and cross-sensory stimuli, and excitatory involved in speech perception (Molholm et al., 2013) that is projections among the unisensory regions of the model were also frequently implicated in audiovisual multisensory subjected to Hebbian rules of potentiation and depression. processing (Bolognini et al., 2009). This region must be With such a model, we also analyzed the acquisition of adult MSI abilities in ASD children, and we were able to explain considered in the context of its feedforward inputs from the delayed maturation as result of a lower level of auditory and visual cortices. Converging evidence reveals multisensory exposures during early phases of life. that MSI occurs at very early stages of cortical processing and in sensory cortical regions, although the functional role Keywords: Autism Spectrum Disorder (ASD); Neural Networks; Hebbian Learning Rules; Multisensory Speech of early MSI (at the onset of cortical sensory processing in Integration; McGurk Effect some cases; Molholm et al., 2002) remains unknown. Several experimental data pointed out that auditory Introduction speech recognition is relatively mature at 5 to 9 years of age, approaching adult-like performances (e.g., Fallon, The brain’s ability to exploit and integrate sensory Trehub & Schneider, 2000; Kraus, Koch, McGee, Nicol, & information of different modalities is fundamental not just Cunningham, 1999), at ages where multisensory speech for simple detection tasks, but also for more demanding processing is not (Foxe et al, 2015). perceptual-cognitive functions, such as those involved in communication. For example, the intelligibility of speech is 389 Such observations suggest a neural model in which the excitatory and inhibitory connections (intra-area maturation of MSI in speech perception follows from the connections, L in Fig. 1), described by a Mexican hat reinforcement of direct “cross-modal” excitatory disposition, i.e., proximal units excite reciprocally and connections between auditory and visual speech inhibit more distal ones. This disposition produces an representations in unisensory cortices. In this case, it can be “activation bubble” in response to a specific auditory or assumed that connections among unisensory areas are visual input: not only the neural element representing that initially relatively ineffective, but that they strengthen as a individual feature is activated, but also the proximal ones consequence of relevant multisensory experiences through a linked via sufficient lateral excitation. This arrangement can Hebbian learning mechanism. Thus, multisensory have important consequences for the correct perception of experiences would affect only the ability of STS elements phonemes, for instance resulting in the illusory perceptual to detect multisensory stimuli, via a reciprocal phenomena like the well-known McGurk effect (see section reinforcement of unisensory activities when both are active, Results). In this work, lateral intra-area connections are not but it would not produce any additional level of information subject to training, since we assumed that this process took to the STS in case of unisensory stimulation. place earlier in life than the acquisition of MSI. The aim of the present work is to test the feasibility of this model, and its consequences by using a computational model inspired by neurophysiology and based on a previous model implemented to study cortical multisensory interaction (Magosso et al., 2012; Cuppini et al., 2014). In particular, with the model we wish to i) analyze possible mechanisms underlying the maturation of MSI; ii) test the model's ability to reproduce different results concerning speech MSI in terms of accuracy as well whether it produces the well-known McGurk illusion; and iii) provide possible explanations of the neural processing differences that could lead to a slower maturation of MSI in participants with ASD, followed by a full recovery during adolescence. In particular, we describe the training mechanisms implemented to simulate the maturation phase and we test a hypothesis to explain ASD deficits in speech MSI: a different multisensory experience during the maturation process, due to a lack of attention in young children (attentional bias) is responsible for the different maturation Figure 1: Architecture of the network. Each circle in ASD. All the simulated responses are compared with represents a unit. Each region is made of 180 elements. behavioral data present in the literature. Dashed lines represent weigths (Wav, Wva) acquired during a crossmodal training, which simulates associative learning Method between speech sound and gestures. Units in the same The model consists of a multisensory region (STS) of N region are reciprocally connected through lateral synapses multisensory units (N = 180), receiving excitatory (La, Lv and Ls), described by a Mexican Hat function. Units projections from two arrays of N auditory and N visual units in the unisensory regions send excitatory connections (Wsv, (see Fig. 1). Unit response to any input is described with a Wsa) to the corresponding elements in the multisensory area. first order differential equation, which simulates the integrative properties of the cellular membrane, and a Furthermore, units in the auditory and visual regions also steady-state sigmoidal relationship, that simulates the receive an external input (corresponding to a speech sound presence of a lower threshold and an upper saturation for and/or a gesture representation of the presented phoneme). neural activation. The saturation value is set at 1, i.e., all These visual and auditory inputs are described with a outputs are normalized to the maximum. In the following, gaussian function. The central point of the Gaussian the term “activity” is used to denote unit output. function corresponds to a specific speech sound/gesture, and Auditory and visual units are devoted to the processing its amplitude with the stimulus intensity; the standard of information regarding speech sounds and speech gestures deviation accounts for the uncertainty of the stimulus (i.e. lip and face movements; see e.g., Bernstein & representation. In this model, for simplicity the two inputs Liebenthal, 2014), and are topologically organized are described with the same function. To reproduce according to a similarity principle. This means that two experimental variability, the external input had been added similar sounds or lips movements activate proximal neural with a noisy component, taken from a uniform distribution. groups in these areas. The topological organization in these Moreover, since the outside inputs are mediated by long- cortical regions is realized assuming that each unit is range excitatory connections, their temporal aspects are connected with other elements of the same area via lateral described by using a second order kinetics, similar to that commonly adopted to mimic the glutamatergic synaptic 390 response (i.e., an impulse produces a response similar to an visual stimuli and 20% of auditory stimuli alone. During the alpha function, see also Jansen and Rit, 1995). These training phase we used suprathreshold stimuli at their kinetics are characterized by different time constants (a and highest level of efficacy, i.e. stimuli able to excite v for the two modalities) simulating the auditory and visual unisensory units close to the upper saturation, in order to processing in the cortex. speed up the modeling process. Finally, we consider a cross-modal input, computed assuming that units of the two areas could be reciprocally linked via long-range excitatory connections (Wav, Wva, in Fig 1), described by a pure latency, and the same second- order kinetics employed to mimic the temporal aspects of the external inputs. We assume that, in the network’s initial configuration, corresponding to an early period of life, these connections have negligible strength, bur are subject to a training phase (see below) during which the network learns to associate the auditory (speech sounds) and visual (speech gestures) representations of the same phonemes. The third area simulates multisensory units in a cortical region (STS) known to be involved in the phoneme comprehension tasks, and MSI. These units are linked via lateral connections with a Mexican-hat arrangement, Figure 2: Representation of training mechanisms simulating implementing a similarity principle (Ls, in Fig. 1). the associative learning between sounds (Auditory area) and Inputs to the multisensory area were generated by long- gestures (Visual area) of units of speech. In case of range excitatory connections from unisensory regions (Wsv, multisensory stimuli, speech sounds are presented along Wsa): we used a delayed onset (pure latency) and a second- with corresponding lip movements. Thanks to the Hebbian order kinetics to mimic the temporal aspects of these inputs. learning rules, connections among contemporarily active The connections between unisensory and multisensory units are reinforced. Hence, the network learns how to regions were realized with a Gaussian function, assuming associate the auditory and visual representations of the same stronger and more focused connections coming from the speech events, and this knowledge is implemented in the auditory region (Wsa), and more diffuse but weaker synaptic architecture between the unisensory regions. connections coming from the visual area (Wsv). This asymmetric connectivity helps explain the experimental These stimuli were generated through a uniform results present in the literature about the better abilities in distribution of probability. Each stimulus lasted 130 ms, speech identification in case of auditory stimulation, during which, after an initial transient period, the compared with the poor performance in the case of visual connections among visual and auditory representations of inputs. This different representation is assumed being the the same phonemes were crafted by using Hebbian final state of a process of unisensory maps refinement in algorithms of long-term potentiation (LTP) and long-term STS, which takes place in early stages of life. This depression (LTD). In particular, we chose a presynaptic development could be included in future implementations of gating rule, which means that the training algorithm only the model, as an earlier training phase based on the evidence modifies the connections coming from an active unit, and that auditory stimuli are more informative than the visual their strength is modified based on the activity of the representations of words. The feedforward connectivity is postsynaptic units. As an example, if a presynaptic auditory also responsible for the presence of early weak integrative element is active, it reinforces connections targeting a phenomena in the younger ASD group (Fig. 1A, Foxe et simultaneously active visual unit (likely representing the al.). In the present model, neither of these connections are same speech unit), and weakens connections with silent modified during the learning period due to the relative visual elements (likely those coding for different speech stability of representations of unisensory speech features at inputs, see Fig. 2) (see Gerstner, W., & Kistler, 2002). In the ages considered (~7 years of age upward). Finally, the order to establish this correlation, the activity of the output of the STS units is compared with a fixed threshold individual units (both presynaptic and postsynaptic) is to mimic the perceptual ability to correctly identify speech compared with a given threshold, to determine whether the (detection threshold). unit can be considered active or silent. The strengthening and depression processes are also subject to a saturation Training the Network rule: which means that each single connection cannot We simulated a normal training period by presenting overcome a maximum value, nor decrease below zero. thousands (up to 25.000 inputs) of unisensory and Finally, to simulate the delayed developmental processes multisensory speech representations to the network, to taking place in ASD children, we trained and tested the mimic a normal experience with speech stimuli: specifically network by using lower multisensory experiences, precisely we trained the network with 80% of congruent auditory and 20% multisensory stimuli plus 80% auditory stimuli. 391 Results Developmental process and audio-visual speech A first set of simulations was performed to evaluate the recognition network’s ability to correctly identify speech before the The model in its initial state was repeatedly stimulated model had been exposed to training (Fig. 3). Already mature with modality-specific and cross-modal inputs (see section unisensory maps in auditory and visual regions were Training) in order to simulate the experience of a child with supposed in this model, as described in previous section. different sensory representations of phonemes. The weights of the inter-area projections among unisensory elements in the visual and auditory regions adjusted according to Hebbian dynamics. We tested MSI in the final “adult-like” configuration and throughout the developmental process, using the same testing paradigm used to evaluate the MSI behavior in the immature phase. Figure 3: Average word recognition performance (% correct) before training. Panel A reports the percentage correct speech recognition (y-axes) in case of auditory stimulation (dashed line), or multisensory stimulation (solid line). These data represent the mean of correct recognitions over 3600 different presentations for each level of stimulus efficacy (reported in the x-axes). A correct recognition has been computed comparing the activity elicited in a unit in the STS region, coding for a specific phoneme, with a threshold (fixed at the 30% of its maximum value). Panel B reports the Multisensory Speech Integration (MSI) abilities of the network, computed as the difference between the percentage of correct detections in case of crossmodal stimulations and its counterpart in case of auditory stimuli. In this phase, representations of speech in the two unisensory regions are independently activated by the two modality-specific external stimuli, and do not interact through direct long-range excitatory projections between the unisensory cortical regions, which are still ineffective. Figure 4: MSI acquisition with a multisensory experience of Hence, they independently stimulate the corresponding units 20% during the training (red lines) compared with the in STS region. As shown in Fig. 3, in this initial condition, normal multisensory experience (blue lines). an effective auditory stimulus alone is sufficient to produce a high percentage of correct speech sound identifications, as One possible explanation for reduced MSI in ASD is that in mature adult-like behaviour. If the auditory stimulus is learning is less effective in this group. A possible coupled with a simultaneous visual representation of the explanation tested here is that these individuals experience same phoneme, the network shows some benefit, although fewer multisensory exposures, possibly due to how attention this is relatively low and no greater than 20% MSI gain over is allocated (e.g., suppression of unattended signals; all stimuli and levels of efficacy. selectively focusing on one sensory modality at a time; not So, the network in its initial stage is characterized by: i) looking at faces consistently). We therefore tested the mature abilities in speech-recognition tasks in case of impact of percentage of multisensory versus unisensory auditory-alone stimulation, but ii) poor multisensory exposures on model performance on the maturation of MSI. integration (see Fig. 3). These results are in agreement with Fig. 4 reports the weight maturation (left panel) and MSI what one would expect prior to significant training, and abilities (right panel) at different epochs, for a training indeed are well aligned with what we see in our data in phase in which the network was exposed to a sensory which younger children show relatively immature ability to training with just 20% of multisensory stimuli. benefit from MSI, whereas auditory speech recognition is Even with such a poor multisensory experience, the significantly closer to mature performance levels. network can reach “TD-like” behaviour in terms of MSI as shown in Fig. 4, although this maturation requires 15,000 392 training epochs. This result suggests that multisensory different perceptual behaviors based on words integration in the model strongly depends on connections representation as a collection of phonetic features in a from the visual to the auditory region. topographically organized feature space. Although the previous computational efforts simulated Simulation of the McGurk effect experimental data quite well, none of them was able to An important consequence of training in our model is that explain the maturation of MSI in speech perception or the the audio-visual inference becomes stronger after training, different developmental trajectory for ASD, or how these because of connection-weight reinforcement among capabilities are instantiated in the circuit. unisensory areas. This change have important consequences The present model, in its mature architecture, simulates in the development of audio-visual illusions. Since many experimental findings present in literature regarding unisensory areas in our model code for speech, a typical speech MSI. From this point of view, the fundamental illusion consists in the well-known McGurk effect (McGurk assumption is that the adult configuration implements a two- & McDonald, 1976). In this illusion, incongruent auditory step cross-modal integration: the first at the level of speech is dubbed onto visual speech and the resulting unisensory areas, mediated by the cross-modal connections auditory speech percept corresponds to a fusion of the between visual and auditory regions; the second at the level auditory and visual speech stimuli, or to the visual speech of the multimodal area, due to the presence of convergent stimulus, but not to the veridical auditory speech stimulus. feedforward connections. With this model, we reproduced We performed an additional set of simulations with the the improvement in correct phoneme recognition in model (both in the mature and immature configurations) to audiovisual vs auditory conditions at different signal-to- reproduce a McGurk-type situation. Specifically, we noise levels (Foxe et al., 2015); and, we simulated the main presented mismatched (at four-position distance) auditory- aspects of the McGurk effect. visual speech to the network and analyzed the activities A second important aspect of our study is the capacity to elicited in all areas. We say that the McGurk effect is mimic and to understand the developmental differences evident when the detected phoneme (computed as the between TD subjects and ASD children regard the cross- barycenter of activity in the multisensory region) is different modal abilities observed with age. In particular, the model from that used in the auditory input. The network in the explains results of a recent study by Foxe et al. (2015), immature configuration is characterized by limited visual using two main assumptions. First, the feedforward influence on the speech percept. Therefore, the activity in connections from unisensory areas to the multisensory area the auditory region is almost unaffected by the visual are already mature in the early age (here, this corresponds to stimulus. In this case, the auditory modality plays the the condition of the untrained network) and the auditory dominant role in guiding speech perception. In the 42.5% of feedforward connections are stronger than the visual ones. presentations, the model identifies the auditory input Second, the cross-modal connections between unimodal correctly, while the McGurk effect is present less than 30% areas are created during the development, under the pressure of the time. In the remaining 27.2% of cases, no phoneme of a multimodal environment (i.e., auditory + visual stimuli) reaches the detection threshold. and this process is faster in TD subjects than in ASDs. This After training, the model is much more susceptible to the assumption agrees with the diffuse idea that ASD subjects AV illusion, with responses affected by the visual have a decreased long-range connectivity, and that autism is information on almost 72% of the simulation trials. a functional disconnection syndrome, in which the core of deficit derives from the poor capacity to functionally Discussion connect remote regions of the brain (Melillo & Leisman, 2009). Since the reason for this decreased connectivity is Different computational models have been developed in still unclear, the model tested a possible scenario where a recent years to investigate the general problem of reduced number of cross-modal stimuli (reflecting a reduced multisensory integration in the brain (see Cuppini et al., attention of the subject to the external world), is a likely 2011 and Ursino et al., 2014 as a review). Some of them, in mechanism responsible for the differences in TD and ASD. agreement with several psychophysical and behavioral data, These differences may lead to some testable predictions: are based on a Bayesian approach (Anastasio et al., 2000; from the results about the training phase, when can expect Knill and Pouget, 2004; Körding et al., 2007). Others that ASD children trained with a high percentage of cross- assume that integration is an emergent property based on modal stimuli, could exhibit a normal or at least a quicker network dynamics (Patton and Anastasio, 2003; Ursino et MSI maturation. A second prediction is that as a al., 2009). Finally some models have been realized to deal consequence of poor cross-modal connections among with the problem of multisensory integration in semantic unisensory areas, young individual with ASDs have a less memory and lexical aspects (Rogers et al., 2004; Ursino et evident McGurk effect, but at the end of the developmental al., 2010, 2015). phase, this illusion becomes comparable in the two classes. Concerning the specific problem of speech recognition, The first prediction is still to be tested; the second is Ma et al (Ma et al., 2009) implemented a Bayesian model of supported by some experimental results in the literature, in optimal cue integration able to explain visual influence on particular by comparing data across different studies. auditory perception in a noisy environment. They explained 393 However, it deserves a deeper investigation through a Körding, K.P., Beierholm, U., Ma, W.J., Quartz, S., single, ideally longitudinal study. Tenenbaum, J.B., & Shams, L. (2007). Causal inference Future developments of this model may include a more in multisensory perception. PLoS ONE 2 (9), e943. detailed and biologically realistic description of the Kraus, N., Koch, D.B., McGee, T.J., Nicol, T.G., & unisensory areas, and the inclusion of further regions to Cunningham, J. (1999). Speech-Sound Discrimination in simulate the role in MSI and speech perception played by School-Age Children. Psychophysical and subcortical structures, like the thalamus and basal ganglia. Neurophysiologic Measures. Journal of Speech, Language, and Hearing Research, 42(5), 1042-1060. References Ma, W.J., Zhou, X., Ross, L.A., Foxe, J.J., & Parra, L.C. Anastasio, T.J., Patton, P.E., & Belkacem-Boussaid, K. (2009). Lip-reading aids word recognition most in (2000) Using Bayes rule to model multisensory moderate noise: a Bayesian explanation using high- enhancement in the superior colliculus. Neural dimensional feature space. PLoS One, 4(3), e4638. Computation, 12, 1165–1187 Magosso, E., Cuppini, C., & Ursino, M. (2012). A neural Bernstein, L.E., & Liebenthal, E. (2014). Neural pathways network model of ventriloquism effect and aftereffect. for visual speech perception. Frontiers in neuroscience, 8. PloS one, 7(8), e42503. de Boer-Schellekens, L., Keetels, M., Eussen, M., & McGurk H, MacDonald J. (1976). Hearing lips and seeing Vroomen, J. (2013). No evidence for impaired voices. Science, 264, 746-8. multisensory integration of low-level audiovisual stimuli Melillo, R., & Leisman, G. (2009). Autistic spectrum in adolescents and young adults with autism spectrum disorders as functional disconnection syndrome. Reviews disorders. Neuropsychologia, 51(14), 3004-3013. in the Neurosciences, 20(2), 111-132. Bolognini, N., Miniussi, C., Savazzi, S., Bricolo, E., & Molholm, S., Ritter, W., Murray, M.M., Javitt, D.C., Maravita, A. (2009). TMS modulation of visual and Schroeder, C. E., & Foxe, J. J. (2002). Multisensory auditory processing in the posterior parietal cortex. auditory–visual interactions during early sensory Experimental brain research, 195(4), 509-517. processing in humans: a high-density electrical mapping Brandwein, A.B., Foxe, J.J., Russo, N.N., Altschuler, T.S., study. Cognitive brain research, 14(1), 115-128. Gomes, H., & Molholm, S. (2011). The development of Molholm, S., Mercier, M.R., Liebenthal, E., Schwartz, T.H., audiovisual multisensory integration across childhood and Ritter, W., Foxe, J.J., & De Sanctis, P. (2014). Mapping early adolescence: a high-density electrical mapping phonemic processing zones along human perisylvian study. Cerebral Cortex, 21(5), 1042-1055. cortex: an electro-corticographic investigation. Brain Cuppini, C., Magosso, E., Bolognini, N., Vallar, G., & Structure and Function, 219, 1369-1383. Ursino, M. (2014). A neurocomputational analysis of the Nath, A.R., & Beauchamp, M.S. (2012). A neural basis for sound-induced flash illusion. NeuroImage, 92, 248-266. interindividual differences in the McGurk effect, a Cuppini, C., Magosso, E., & Ursino, M. (2011). multisensory speech illusion. Neuroimage, 59, 781-787. Organization, maturation, and plasticity of multisensory Patton, P.E., & Anastasio, T.J. (2003). Modeling cross- integration: insights from computational modeling modal enhancement and modality-specific suppression in studies. Frontiers in psychology, 2. multisensory neurons. Neural computation, 15, 783-810. Fallon, M., Trehub, S.E., & Schneider, B.A. (2000). Rogers, T.T., Lambon Ralph, M.A., Garrard, P., Bozeat, S., Children’s perception of speech in multitalker babble. The McClelland, J.L., Hodges, J.R., and Patterson, K. (2004). Journal of the Acoustical Society of America, 108(6), The structure and deterioration of semantic memory: A 3023-3029. neuropsychological and computational investigation. Foxe, J.J., Molholm, S., Del Bene, V.A., Frey, H. P., Russo, Psychological Review, 111, 205-235. N.N., Blanco, D., Saint-Amour, D., & Ross, L.A. (2015). Ursino, M., Cuppini, C., & Magosso, E. (2010). A Severe multisensory speech integration deficits in high- computational model of the lexical semantic system based functioning school-aged children with autism spectrum on a grounded cognition approach. Frontiers in disorder (ASD) and their resolution during early Psychology, 1, 221. adolescence. Cerebral Cortex, 25(2), 298-312. Ursino, M., Cuppini, C., & Magosso, E. (2014). Gerstner, W., & Kistler, W.M. (2002). Mathematical Neurocomputational approaches to modelling formulations of Hebbian learning. Biological cybernetics, multisensory integration in the brain: A review. Neural 87(5-6), 404-415. Networks, 60, 141-165. Jansen, B.H., & Rit, V.G. (1995). Electroencephalogram Ursino, M., Cuppini, C., & Magosso, E. (2015). A neural and visual evoked potential generation in a mathematical network for learning the meaning of objects and words model of coupled cortical columns. Biological from a featural representation. Neural Networks, 63, 234- cybernetics, 73(4), 357-366. 253. Knill, D.C., & Pouget, A. (2004). The Bayesian brain: the Ursino, M., Cuppini, C., Magosso, E., Serino, A., & Di role of uncertainty in neural coding and computation. Pellegrino, G. (2009). Multisensory integration in the Trends in Neurosciences, 27 (12), 712–719. superior colliculus: a neural network model. Journal of computational neuroscience, 26(1), 55-73. 394