Using Bayes' rule to model multisensory enhancement in the superior colliculus

Multisensory Concept Learning Framework Based on Spiking Neural Networks

Concept learning highly depends on multisensory integration. In this study, we propose a multisensory concept learning framework based on brain-inspired spiking neural networks to create integrated vectors relying on the concept's perceptual strength of auditory, gustatory, haptic, olfactory, and visual. With different assumptions, two paradigms: Independent Merge (IM) and Associate Merge (AM) are designed in the framework. For testing, we employed eight distinct neural models and three multisensory representation datasets. The experiments show that integrated vectors are closer to human beings than the non-integrated ones. Furthermore, we systematically analyze the similarities and differences between IM and AM paradigms and validate the generality of our framework.

Cross-modal perceptual enhancement of unisensory targets is uni-directional and does not affect temporal expectations

Temporal structures in the environment can shape temporal expectations (TE); and previous studies demonstrated that TEs interact with multisensory interplay (MSI) when multisensory stimuli are presented synchronously. Here, we tested whether other types of MSI - evoked by asynchronous yet temporally flanking irrelevant stimuli - result in similar performance patterns. To this end, we presented sequences of 12 stimuli (10 Hz) which consisted of auditory (A), visual (V) or alternating auditory-visual stimuli (e.g. A-V-A-V-…) with either auditory or visual targets (Exp. 1). Participants discriminated target frequencies (auditory pitch or visual spatial frequency) embedded in these sequences. To test effects of TE, the proportion of early and late temporal target positions was manipulated run-wise. Performance for unisensory targets was affected by temporally flanking distractors, with auditory temporal flankers selectively improving visual target perception (Exp. 1). However, no effect of temporal expectation was observed. Control experiments (Exp. 2-3) tested whether this lack of TE effect was due to the higher presentation frequency in Exp. 1 relative to previous experiments. Importantly, even at higher stimulation frequencies redundant multisensory targets (Exp. 2-3) reliably modulated TEs. Together, our results indicate that visual target detection was enhanced by MSI. However, this cross-modal enhancement - in contrast to the redundant target effect - was still insufficient to generate TEs. We posit that unisensory target representations were either instable or insufficient for the generation of TEs while less demanding MSI still occurred; highlighting the need for robust stimulus representations when generating temporal expectations.

A Biologically Motivated, Proto-Object-Based Audiovisual Saliency Model

The natural environment and our interaction with it are essentially multisensory, where we may deploy visual, tactile and/or auditory senses to perceive, learn and interact with our environment. Our objective in this study is to develop a scene analysis algorithm using multisensory information, specifically vision and audio. We develop a proto-object-based audiovisual saliency map (AVSM) for the analysis of dynamic natural scenes. A specialized audiovisual camera with 360∘ field of view, capable of locating sound direction, is used to collect spatiotemporally aligned audiovisual data. We demonstrate that the performance of a proto-object-based audiovisual saliency map in detecting and localizing salient objects/events is in agreement with human judgment. In addition, the proto-object-based AVSM that we compute as a linear combination of visual and auditory feature conspicuity maps captures a higher number of valid salient events compared to unisensory saliency maps. Such an algorithm can be useful in surveillance, robotic navigation, video compression and related applications.

Audio-visual spatial alignment improves integration in the presence of a competing audio-visual stimulus

In order to parse the world around us, we must constantly determine which sensory inputs arise from the same physical source and should therefore be perceptually integrated. Temporal coherence between auditory and visual stimuli drives audio-visual (AV) integration, but the role played by AV spatial alignment is less well understood. Here, we manipulated AV spatial alignment and collected electroencephalography (EEG) data while human subjects performed a free-field variant of the "pip and pop" AV search task. In this paradigm, visual search is aided by a spatially uninformative auditory tone, the onsets of which are synchronized to changes in the visual target. In Experiment 1, tones were either spatially aligned or spatially misaligned with the visual display. Regardless of AV spatial alignment, we replicated the key pip and pop result of improved AV search times. Mirroring the behavioral results, we found an enhancement of early event-related potentials (ERPs), particularly the auditory N1 component, in both AV conditions. We demonstrate that both top-down and bottom-up attention contribute to these N1 enhancements. In Experiment 2, we tested whether spatial alignment influences AV integration in a more challenging context with competing multisensory stimuli. An AV foil was added that visually resembled the target and was synchronized to its own stream of synchronous tones. The visual components of the AV target and AV foil occurred in opposite hemifields; the two auditory components were also in opposite hemifields and were either spatially aligned or spatially misaligned with the visual components to which they were synchronized. Search was fastest when the auditory and visual components of the AV target (and the foil) were spatially aligned. Attention modulated ERPs in both spatial conditions, but importantly, the scalp topography of early evoked responses shifted only when stimulus components were spatially aligned, signaling the recruitment of different neural generators likely related to multisensory integration. These results suggest that AV integration depends on AV spatial alignment when stimuli in both modalities compete for selective integration, a common scenario in real-world perception.

Tutorial Review of Bio-Inspired Approaches to Robotic Manipulation for Space Debris Salvage

We present a comprehensive tutorial review that explores the application of bio-inspired approaches to robot control systems for grappling and manipulating a wide range of space debris targets. Current robot manipulator control systems exploit limited techniques which can be supplemented by additional bio-inspired methods to provide a robust suite of robot manipulation technologies. In doing so, we review bio-inspired control methods because this will be the key to enabling such capabilities. In particular, force feedback control may be supplemented with predictive forward models and software emulation of viscoelastic preflexive joint behaviour. This models human manipulation capabilities as implemented by the cerebellum and muscles/joints respectively. In effect, we are proposing a three-level control strategy based on biomimetic forward models for predictive estimation, traditional feedback control and biomimetic muscle-like preflexes. We place emphasis on bio-inspired forward modelling suggesting that all roads lead to this solution for robust and adaptive manipulator control. This promises robust and adaptive manipulation for complex tasks in salvaging space debris.

Contextual Integration in Cortical and Convolutional Neural Networks

It has been suggested that neurons can represent sensory input using probability distributions and neural circuits can perform probabilistic inference. Lateral connections between neurons have been shown to have non-random connectivity and modulate responses to stimuli within the classical receptive field. Large-scale efforts mapping local cortical connectivity describe cell type specific connections from inhibitory neurons and like-to-like connectivity between excitatory neurons. To relate the observed connectivity to computations, we propose a neuronal network model that approximates Bayesian inference of the probability of different features being present at different image locations. We show that the lateral connections between excitatory neurons in a circuit implementing contextual integration in this should depend on correlations between unit activities, minus a global inhibitory drive. The model naturally suggests the need for two types of inhibitory gates (normalization, surround inhibition). First, using natural scene statistics and classical receptive fields corresponding to simple cells parameterized with data from mouse primary visual cortex, we show that the predicted connectivity qualitatively matches with that measured in mouse cortex: neurons with similar orientation tuning have stronger connectivity, and both excitatory and inhibitory connectivity have a modest spatial extent, comparable to that observed in mouse visual cortex. We incorporate lateral connections learned using this model into convolutional neural networks. Features are defined by supervised learning on the task, and the lateral connections provide an unsupervised learning of feature context in multiple layers. Since the lateral connections provide contextual information when the feedforward input is locally corrupted, we show that incorporating such lateral connections into convolutional neural networks makes them more robust to noise and leads to better performance on noisy versions of the MNIST dataset. Decomposing the predicted lateral connectivity matrices into low-rank and sparse components introduces additional cell types into these networks. We explore effects of cell-type specific perturbations on network computation. Our framework can potentially be applied to networks trained on other tasks, with the learned lateral connections aiding computations implemented by feedforward connections when the input is unreliable and demonstrate the potential usefulness of combining supervised and unsupervised learning techniques in real-world vision tasks.

Interhemispheric visual competition after multisensory reversal of hemianopia

Unilateral lesions of visual cortex have the secondary consequence of suppressing visual circuits in the midbrain superior colliculus (SC), collectively producing blindness in contralesional space (“hemianopia”). Recent studies have demonstrated that SC visual responses and contralesional vision can be reinstated by a non‐invasive multisensory training procedure in which spatiotemporally concordant visual‐auditory pairs are repeatedly presented within the blind hemifield. Despite this recovery of visual responsiveness, the loss of visual cortex was expected to result in permanent deficits in that hemifield, especially when visual events in both hemifields compete for attention and access to the brain's visuomotor circuitry. This was evaluated in the present study in a visual choice paradigm in which the two visual hemifields of recovered cats were simultaneously stimulated with equally valent visual targets. Surprisingly, the expected disparity was not found, and some animals even preferred stimuli presented in the previously blind hemifield. This preference persisted across multiple stimulus intensity levels and there was no indication that animals were less aware of cues in the previously blind hemifield than in its spared counterpart. Furthermore, when auditory cues were combined with visual cues, the enhanced performance they produced on a visual task was no greater in the normal than in the previously blind hemifield. These observations suggest that the multisensory rehabilitation paradigm revealed greater inherent visual information processing potential in the previously blind hemifield than was believed possible given the loss of visual cortex.

Probabilistic Representation in Human Visual Cortex Reflects Uncertainty in Serial Decisions

How does the brain represent the reliability of its sensory evidence? Here, we test whether sensory uncertainty is encoded in cortical population activity as the width of a probability distribution, a hypothesis that lies at the heart of Bayesian models of neural coding. We probe the neural representation of uncertainty by capitalizing on a well-known behavioral bias called serial dependence. Human observers of either sex reported the orientation of stimuli presented in sequence, while activity in visual cortex was measured with fMRI. We decoded probability distributions from population-level activity and found that serial dependence effects in behavior are consistent with a statistically advantageous sensory integration strategy, in which uncertain sensory information is given less weight. More fundamentally, our results suggest that probability distributions decoded from human visual cortex reflect the sensory uncertainty that observers rely on in their decisions, providing critical evidence for Bayesian theories of perception.SIGNIFICANCE STATEMENT Virtually any decision that people make is based on uncertain and incomplete information. Although uncertainty plays a major role in decision-making, we have but a nascent understanding of its neural basis. Here, we probe the neural code of uncertainty by capitalizing on a well-known perceptual illusion. We developed a computational model to explain the illusion, and tested it in behavioral and neuroimaging experiments. This revealed that the illusion is not a mistake of perception, but rather reflects a rational decision under uncertainty. No less important, we discovered that the uncertainty that people use in this decision is represented in brain activity as the width of a probability distribution, providing critical evidence for current Bayesian theories of decision-making.

An Extreme Value Theory Model of Cross-Modal Sensory Information Integration in Modulation of Vertebrate Visual System Functions

We propose a computational model of vision that describes the integration of cross-modal sensory information between the olfactory and visual systems in zebrafish based on the principles of the statistical extreme value theory. The integration of olfacto-retinal information is mediated by the centrifugal pathway that originates from the olfactory bulb and terminates in the neural retina. Motivation for using extreme value theory stems from physiological evidence suggesting that extremes and not the mean of the cell responses direct cellular activity in the vertebrate brain. We argue that the visual system, as measured by retinal ganglion cell responses in spikes/sec, follows an extreme value process for sensory integration and the increase in visual sensitivity from the olfactory input can be better modeled using extreme value distributions. As zebrafish maintains high evolutionary proximity to mammals, our model can be extended to other vertebrates as well.

Multimodal Object Classification Models Inspired by Multisensory Integration in the Brain

Two multimodal classification models aimed at enhancing object classification through the integration of semantically congruent unimodal stimuli are introduced. The feature-integrating model, inspired by multisensory integration in the subcortical superior colliculus, combines unimodal features which are subsequently classified by a multimodal classifier. The decision-integrating model, inspired by integration in primary cortical areas, classifies unimodal stimuli independently using unimodal classifiers and classifies the combined decisions using a multimodal classifier. The multimodal classifier models are implemented using multilayer perceptrons and multivariate statistical classifiers. Experiments involving the classification of noisy and attenuated auditory and visual representations of ten digits are designed to demonstrate the properties of the multimodal classifiers and to compare the performances of multimodal and unimodal classifiers. The experimental results show that the multimodal classification systems exhibit an important aspect of the "inverse effectiveness principle" by yielding significantly higher classification accuracies when compared with those of the unimodal classifiers. Furthermore, the flexibility offered by the generalized models enables the simulations and evaluations of various combinations of multimodal stimuli and classifiers under varying uncertainty conditions.

Cross-Modal Competition: The Default Computation for Multisensory Processing

Mature multisensory superior colliculus (SC) neurons integrate information across the senses to enhance their responses to spatiotemporally congruent cross-modal stimuli. The development of this neurotypic feature of SC neurons requires experience with cross-modal cues. In the absence of such experience the response of an SC neuron to congruent cross-modal cues is no more robust than its response to the most effective component cue. This "default" or "naive" state is believed to be one in which cross-modal signals do not interact. The present results challenge this characterization by identifying interactions between visual-auditory signals in male and female cats reared without visual-auditory experience. By manipulating the relative effectiveness of the visual and auditory cross-modal cues that were presented to each of these naive neurons, an active competition between cross-modal signals was revealed. Although contrary to current expectations, this result is explained by a neuro-computational model in which the default interaction is mutual inhibition. These findings suggest that multisensory neurons at all maturational stages are capable of some form of multisensory integration, and use experience with cross-modal stimuli to transition from their initial state of competition to their mature state of cooperation. By doing so, they develop the ability to enhance the physiological salience of cross-modal events thereby increasing their impact on the sensorimotor circuitry of the SC, and the likelihood that biologically significant events will elicit SC-mediated overt behaviors.SIGNIFICANCE STATEMENT The present results demonstrate that the default mode of multisensory processing in the superior colliculus is competition, not non-integration as previously characterized. A neuro-computational model explains how these competitive dynamics can be implemented via mutual inhibition, and how this default mode is superseded by the emergence of cooperative interactions during development.

Development of the Mechanisms Governing Midbrain Multisensory Integration

The ability to integrate information across multiple senses enhances the brain's ability to detect, localize, and identify external events. This process has been well documented in single neurons in the superior colliculus (SC), which synthesize concordant combinations of visual, auditory, and/or somatosensory signals to enhance the vigor of their responses. This increases the physiological salience of crossmodal events and, in turn, the speed and accuracy of SC-mediated behavioral responses to them. However, this capability is not an innate feature of the circuit and only develops postnatally after the animal acquires sufficient experience with covariant crossmodal events to form links between their modality-specific components. Of critical importance in this process are tectopetal influences from association cortex. Recent findings suggest that, despite its intuitive appeal, a simple generic associative rule cannot explain how this circuit develops its ability to integrate those crossmodal inputs to produce enhanced multisensory responses. The present neurocomputational model explains how this development can be understood as a transition from a default state in which crossmodal SC inputs interact competitively to one in which they interact cooperatively. Crucial to this transition is the operation of a learning rule requiring coactivation among tectopetal afferents for engagement. The model successfully replicates findings of multisensory development in normal cats and cats of either sex reared with special experience. In doing so, it explains how the cortico-SC projections can use crossmodal experience to craft the multisensory integration capabilities of the SC and adapt them to the environment in which they will be used.SIGNIFICANCE STATEMENT The brain's remarkable ability to integrate information across the senses is not present at birth, but typically develops in early life as experience with crossmodal cues is acquired. Recent empirical findings suggest that the mechanisms supporting this development must be more complex than previously believed. The present work integrates these data with what is already known about the underlying circuit in the midbrain to create and test a mechanistic model of multisensory development. This model represents a novel and comprehensive framework that explains how midbrain circuits acquire multisensory experience and reveals how disruptions in this neurotypic developmental trajectory yield divergent outcomes that will affect the multisensory processing capabilities of the mature brain.

A Computational Analysis of Neural Mechanisms Underlying the Maturation of Multisensory Speech Integration in Neurotypical Children and Those on the Autism Spectrum

Failure to appropriately develop multisensory integration (MSI) of audiovisual speech may affect a child's ability to attain optimal communication. Studies have shown protracted development of MSI into late-childhood and identified deficits in MSI in children with an autism spectrum disorder (ASD). Currently, the neural basis of acquisition of this ability is not well understood. Here, we developed a computational model informed by neurophysiology to analyze possible mechanisms underlying MSI maturation, and its delayed development in ASD. The model posits that strengthening of feedforward and cross-sensory connections, responsible for the alignment of auditory and visual speech sound representations in posterior superior temporal gyrus/sulcus, can explain behavioral data on the acquisition of MSI. This was simulated by a training phase during which the network was exposed to unisensory and multisensory stimuli, and projections were crafted by Hebbian rules of potentiation and depression. In its mature architecture, the network also reproduced the well-known multisensory McGurk speech effect. Deficits in audiovisual speech perception in ASD were well accounted for by fewer multisensory exposures, compatible with a lack of attention, but not by reduced synaptic connectivity or synaptic plasticity.

Effects of Aging in Multisensory Integration: A Systematic Review

Multisensory integration (MSI) is the integration by the brain of environmental information acquired through more than one sense. Accurate MSI has been shown to be a key component of successful aging and to be crucial for processes underlying activities of daily living (ADLs). Problems in MSI could prevent older adults (OA) to age in place and live independently. However, there is a need to know how to assess changes in MSI in individuals. This systematic review provides an overview of tests assessing the effect of age on MSI in the healthy elderly population (aged 60 years and older). A literature search was done in Scopus. Articles from the earliest records available to January 20, 2016, were eligible for inclusion if assessing effects of aging on MSI in the healthy elderly population compared to younger adults (YA). These articles were rated for risk of bias with the Newcastle-Ottawa quality assessment. Out of 307 identified research articles, 49 articles were included for final review, describing 69 tests. The review indicated that OA maximize the use of multiple sources of information in comparison to YA (20 studies). In tasks that require more cognitive function, or when participants need to adapt rapidly to a situation, or when a dual task is added to the experiment, OA have problems selecting and integrating information properly as compared to YA (19 studies). Additionally, irrelevant or wrong information (i.e., distractors) has a greater impact on OA than on YA (21 studies). OA failing to weigh sensory information properly, has not been described in previous reviews. Anatomical changes (i.e., reduction of brain volume and differences of brain areas' recruitment) and information processing changes (i.e., general cognitive slowing, inverse effectiveness, larger time window of integration, deficits in attentional control and increased noise at baseline) can only partly explain the differences between OA and YA regarding MSI. Since we have an interest in successful aging and early detection of MSI issues in the elderly population, the identified tests form a good starting point to develop a clinically useful toolkit to assess MSI in healthy OA.

A Model of the Neural Mechanisms Underlying Multisensory Integration in the Superior Colliculus

Much of the information about multisensory integration is derived from studies of the cat superior colliculus (SC), a midbrain structure involved in orientation behaviors. This integration is apparent in the enhanced responses of SC neurons to cross-modal stimuli, responses that exceed those to any of the modality-specific component stimuli. The simplest model of multisensory integration is one in which the SC neuron simply sums its various sensory inputs. However, a number of empirical findings reveal the inadequacy of such a model; for example, the finding that deactivation of cortico-collicular inputs eliminates the enhanced response to a cross-modal stimulus without eliminating responses to the modality-specific component stimuli. These and other empirical findings inform a computational model that accounts for all of the most fundamental aspects of SC multisensory integration. The model is presented in two forms: an algebraic form that conveys the essential insights, and a compartmental form that represents the neuronal computations in a more biologically realistic way.

Lack of multisensory integration in hemianopia: no influence of visual stimuli on aurally guided saccades to the blind hemifield

In patients with visual hemifield defects residual visual functions may be present, a phenomenon called blindsight. The superior colliculus (SC) is part of the spared pathway that is considered to be responsible for this phenomenon. Given that the SC processes input from different modalities and is involved in the programming of saccadic eye movements, the aim of the present study was to examine whether multimodal integration can modulate oculomotor competition in the damaged hemifield. We conducted two experiments with eight patients who had visual field defects due to lesions that affected the retinogeniculate pathway but spared the retinotectal direct SC pathway. They had to make saccades to an auditory target that was presented alone or in combination with a visual stimulus. The visual stimulus could either be spatially coincident with the auditory target (possibly enhancing the auditory target signal), or spatially disparate to the auditory target (possibly competing with the auditory tar-get signal). For each patient we compared the saccade endpoint deviation in these two bi-modal conditions with the endpoint deviation in the unimodal condition (auditory target alone). In all seven hemianopic patients, saccade accuracy was affected only by visual stimuli in the intact, but not in the blind visual field. In one patient with a more limited quadrantano-pia, a facilitation effect of the spatially coincident visual stimulus was observed. We conclude that our results show that multisensory integration is infrequent in the blind field of patients with hemianopia.

Neural coding of uncertainty and probability

Organisms must act in the face of sensory, motor, and reward uncertainty stemming from a pandemonium of stochasticity and missing information. In many tasks, organisms can make better decisions if they have at their disposal a representation of the uncertainty associated with task-relevant variables. We formalize this problem using Bayesian decision theory and review recent behavioral and neural evidence that the brain may use knowledge of uncertainty, confidence, and probability.

Predictive coding and multisensory integration: an attentional account of the multisensory mind

Multisensory integration involves a host of different cognitive processes, occurring at different stages of sensory processing. Here I argue that, despite recent insights suggesting that multisensory interactions can occur at very early latencies, the actual integration of individual sensory traces into an internally consistent mental representation is dependent on both top–down and bottom–up processes. Moreover, I argue that this integration is not limited to just sensory inputs, but that internal cognitive processes also shape the resulting mental representation. Studies showing that memory recall is affected by the initial multisensory context in which the stimuli were presented will be discussed, as well as several studies showing that mental imagery can affect multisensory illusions. This empirical evidence will be discussed from a predictive coding perspective, in which a central top–down attentional process is proposed to play a central role in coordinating the integration of all these inputs into a coherent mental representation.

Vision-guided robot hearing

Natural human–robot interaction (HRI) in complex and unpredictable environments is important with many potential applications. While vision-based HRI has been thoroughly investigated, robot hearing and audio-based HRI are emerging research topics in robotics. In typical real-world scenarios, humans are at some distance from the robot and, hence, the sensory (microphone) data are strongly impaired by background noise, reverberations and competing auditory sources. In this context, the detection and localization of speakers plays a key role that enables several tasks, such as improving the signal-to-noise ratio for speech recognition, speaker recognition, speaker tracking, etc. In this paper we address the problem of how to detect and localize people that are both seen and heard. We introduce a hybrid deterministic/probabilistic model. The deterministic component allows us to map 3D visual data onto a 1D auditory space. The probabilistic component of the model enables the visual features to guide the grouping of the auditory features in order to form audiovisual (AV) objects. The proposed model and the associated algorithms are implemented in real-time (17 FPS) using a stereoscopic camera pair and two microphones embedded into the head of the humanoid robot NAO. We perform experiments with (i) synthetic data, (ii) publicly available data gathered with an audiovisual robotic head, and (iii) data acquired using the NAO robot. The results validate the approach and are an encouragement to investigate how vision and hearing could be further combined for robust HRI.

Color selectivity of the spatial congruency effect: evidence from the focused attention paradigm

The multisensory response enhancement (MRE), occurring when the response to a visual target integrated with a spatially congruent sound is stronger than the response to the visual target alone, is believed to be mediated by the superior colliculus (SC) (Stein & Meredith, 1993). Here, we used a focused attention paradigm to show that the spatial congruency effect occurs with red (SC-effective) but not blue (SC-ineffective) visual stimuli, when presented with spatially congruent sounds. To isolate the chromatic component of SC-ineffective targets and to demonstrate the selectivity of the spatial congruency effect we used the random luminance modulation technique (Experiment 1) and the tritanopic technique (Experiment 2). Our results indicate that the spatial congruency effect does not require the distribution of attention over different sensory modalities and provide correlational evidence that the SC mediates the effect.

Bridging the divide between sensory integration and binding theory: Using a binding-like neural synchronization mechanism to model sensory enhancements during multisensory interactions

Neural information combination problems are ubiquitous in cognitive neuroscience. Two important disciplines, although conceptually similar, take radically different approaches to these problems. Sensory binding theory is largely grounded in synchronization of neurons responding to different aspects of a stimulus, resulting in a coherent percept. Sensory integration focuses more on the influences of the senses on each other and is largely grounded in the study of neurons that respond to more than one sense. It would be desirable to bridge these disciplines, so that insights gleaned from either could be harnessed by the other. To link these two fields, we used a binding-like oscillatory synchronization mechanism to simulate neurons in rattlesnake that are driven by one sense but modulated by another. Mutual excitatory coupling produces synchronized trains of action potentials with enhanced firing rates. The same neural synchronization mechanism models the behavior of a population of cells in cat visual cortex that are modulated by auditory activation. The coupling strength of the synchronizing neurons is crucial to the outcome; a criterion of strong coupling (kept weak enough to avoid seriously distorting action potential amplitude) results in intensity-dependent sensory enhancement-the principle of inverse effectiveness-a key property of sensory integration.

A model of the temporal dynamics of multisensory enhancement

The senses transduce different forms of environmental energy, and the brain synthesizes information across them to enhance responses to salient biological events. We hypothesize that the potency of multisensory integration is attributable to the convergence of independent and temporally aligned signals derived from cross-modal stimulus configurations onto multisensory neurons. The temporal profile of multisensory integration in neurons of the deep superior colliculus (SC) is consistent with this hypothesis. The responses of these neurons to visual, auditory, and combinations of visual-auditory stimuli reveal that multisensory integration takes place in real-time; that is, the input signals are integrated as soon as they arrive at the target neuron. Interactions between cross-modal signals may appear to reflect linear or nonlinear computations on a moment-by-moment basis, the aggregate of which determines the net product of multisensory integration. Modeling observations presented here suggest that the early nonlinear components of the temporal profile of multisensory integration can be explained with a simple spiking neuron model, and do not require more sophisticated assumptions about the underlying biology. A transition from nonlinear "super-additive" computation to linear, additive computation can be accomplished via scaled inhibition. The findings provide a set of design constraints for artificial implementations seeking to exploit the basic principles and potency of biological multisensory integration in contexts of sensory substitution or augmentation.

Probabilistic brains: knowns and unknowns

There is strong behavioral and physiological evidence that the brain both represents probability distributions and performs probabilistic inference. Computational neuroscientists have started to shed light on how these probabilistic representations and computations might be implemented in neural circuits. One particularly appealing aspect of these theories is their generality: they can be used to model a wide range of tasks, from sensory processing to high-level cognition. To date, however, these theories have only been applied to very simple tasks. Here we discuss the challenges that will emerge as researchers start focusing their efforts on real-life computations, with a focus on probabilistic learning, structural learning and approximate inference.

Hebbian mechanisms help explain development of multisensory integration in the superior colliculus: a neural network model

The superior colliculus (SC) integrates relevant sensory information (visual, auditory, somatosensory) from several cortical and subcortical structures, to program orientation responses to external events. However, this capacity is not present at birth, and it is acquired only through interactions with cross-modal events during maturation. Mathematical models provide a quantitative framework, valuable in helping to clarify the specific neural mechanisms underlying the maturation of the multisensory integration in the SC. We extended a neural network model of the adult SC (Cuppini et al., Front Integr Neurosci 4:1-15, 2010) to describe the development of this phenomenon starting from an immature state, based on known or suspected anatomy and physiology, in which: (1) AES afferents are present but weak, (2) Responses are driven from non-AES afferents, and (3) The visual inputs have a marginal spatial tuning. Sensory experience was modeled by repeatedly presenting modality-specific and cross-modal stimuli. Synapses in the network were modified by simple Hebbian learning rules. As a consequence of this exposure, (1) Receptive fields shrink and come into spatial register, and (2) SC neurons gained the adult characteristic integrative properties: enhancement, depression, and inverse effectiveness. Importantly, the unique architecture of the model guided the development so that integration became dependent on the relationship between the cortical input and the SC. Manipulations of the statistics of the experience during the development changed the integrative profiles of the neurons, and results matched well with the results of physiological studies.

Organizing probabilistic models of perception

Probability has played a central role in models of perception for more than a century, but a look at probabilistic concepts in the literature raises many questions. Is being Bayesian the same as being optimal? Are recent Bayesian models fundamentally different from classic signal detection theory models? Do findings of near-optimal inference provide evidence that neurons compute with probability distributions? This review aims to disentangle these concepts and to classify empirical evidence accordingly.

Focused attention vs. crossmodal signals paradigm: deriving predictions from the time-window-of-integration model

In the crossmodal signals paradigm (CSP) participants are instructed to respond to a set of stimuli from different modalities, presented more or less simultaneously, as soon as a stimulus from any modality has been detected. In the focused attention paradigm (FAP), on the other hand, responses should only be made to a stimulus from a pre-defined target modality and stimuli from non-target modalities should be ignored. Whichever paradigm is being applied, a typical result is that responses tend to be faster to crossmodal stimuli than to unimodal stimuli, a phenomenon often referred to as “crossmodal interaction.” Here, we investigate predictions of the time-window-of-integration (TWIN) modeling framework previously proposed by the authors. It is shown that TWIN makes specific qualitative and quantitative predictions on how the two paradigms differ with respect to the probability of multisensory integration and the amount of response enhancement, including the effect of stimulus intensity (“inverse effectiveness”). Introducing a decision-theoretic framework for TWIN further allows comparing the two paradigms with respect to the predicted optimal time window size and its dependence on the prior probability that the crossmodal stimulus information refers to the same event. In order to test these predictions, experimental studies that systematically compare crossmodal effects under stimulus conditions that are identical except for the CSP-FAP instruction should be performed in the future.

Impact of response duration on multisensory integration

Multisensory neurons in the superior colliculus (SC) have been shown to have large receptive fields that are heterogeneous in nature. These neurons have the capacity to integrate their different sensory inputs, a process that has been shown to depend on the physical characteristics of the stimuli that are combined (i.e., spatial and temporal relationship and relative effectiveness). Recent work has highlighted the interdependence of these factors in driving multisensory integration, adding a layer of complexity to our understanding of multisensory processes. In the present study our goal was to add to this understanding by characterizing how stimulus location impacts the temporal dynamics of multisensory responses in cat SC neurons. The results illustrate that locations within the spatial receptive fields (SRFs) of these neurons can be divided into those showing short-duration responses and long-duration response profiles. Most importantly, discharge duration appears to be a good determinant of multisensory integration, such that short-duration responses are typically associated with a high magnitude of multisensory integration (i.e., superadditive responses) while long-duration responses are typically associated with low integrative capacity. These results further reinforce the complexity of the integrative features of SC neurons and show that the large SRFs of these neurons are characterized by vastly differing temporal dynamics, dynamics that strongly shape the integrative capacity of these neurons.

A rational analysis of the acquisition of multisensory representations

How do people learn multisensory, or amodal, representations, and what consequences do these representations have for perceptual performance? We address this question by performing a rational analysis of the problem of learning multisensory representations. This analysis makes use of a Bayesian nonparametric model that acquires latent multisensory features that optimally explain the unisensory features arising in individual sensory modalities. The model qualitatively accounts for several important aspects of multisensory perception: (a) it integrates information from multiple sensory sources in such a way that it leads to superior performances in, for example, categorization tasks; (b) its performances suggest that multisensory training leads to better learning than unisensory training, even when testing is conducted in unisensory conditions; (c) its multisensory representations are modality invariant; and (d) it predicts ''missing" sensory representations in modalities when the input to those modalities is absent. Our rational analysis indicates that all of these aspects emerge as part of the optimal solution to the problem of learning to represent complex multisensory environments.

Visual signals bias auditory targets in azimuth and depth

In the psychophysical phenomenon visual bias, an accurately localized irrelevant signal, such as a light, impairs localization of a spatially discrepant target, such as a sound, when the two stimuli are perceived as unified. Many studies have demonstrated visual bias in azimuth, but none have tested directly or found this effect in depth. The current study was able to produce over 90% bias in azimuth and somewhat less (83%) bias in depth. A maximum likelihood estimate can predict bias by the variance in the localization of each unimodal signal in each dimension in space.

Computing an optimal time window of audiovisual integration in focused attention tasks: illustrated by studies on effect of age and prior knowledge

The concept of a "time window of integration" holds that information from different sensory modalities must not be perceived too far apart in time in order to be integrated into a multisensory perceptual event. Empirical estimates of window width differ widely, however, ranging from 40 to 600 ms depending on context and experimental paradigm. Searching for theoretical derivation of window width, Colonius and Diederich (Front Integr Neurosci 2010) developed a decision-theoretic framework using a decision rule that is based on the prior probability of a common source, the likelihood of temporal disparities between the unimodal signals, and the payoff for making right or wrong decisions. Here, this framework is extended to the focused attention task where subjects are asked to respond to signals from a target modality only. Evoking the framework of the time-window-of-integration (TWIN) model, an explicit expression for optimal window width is obtained. The approach is probed on two published focused attention studies. The first is a saccadic reaction time study assessing the efficiency with which multisensory integration varies as a function of aging. Although the window widths for young and older adults differ by nearly 200 ms, presumably due to their different peripheral processing speeds, neither of them deviates significantly from the optimal values. In the second study, head saccadic reactions times to a perfectly aligned audiovisual stimulus pair had been shown to depend on the prior probability of spatial alignment. Intriguingly, they reflected the magnitude of the time-window widths predicted by our decision-theoretic framework, i.e., a larger time window is associated with a higher prior probability.

Organization, maturation, and plasticity of multisensory integration: insights from computational modeling studies

In this paper, we present two neural network models – devoted to two specific and widely investigated aspects of multisensory integration – in order to evidence the potentialities of computational models to gain insight into the neural mechanisms underlying organization, development, and plasticity of multisensory integration in the brain. The first model considers visual–auditory interaction in a midbrain structure named superior colliculus (SC). The model is able to reproduce and explain the main physiological features of multisensory integration in SC neurons and to describe how SC integrative capability – not present at birth – develops gradually during postnatal life depending on sensory experience with cross-modal stimuli. The second model tackles the problem of how tactile stimuli on a body part and visual (or auditory) stimuli close to the same body part are integrated in multimodal parietal neurons to form the perception of peripersonal (i.e., near) space. The model investigates how the extension of peripersonal space – where multimodal integration occurs – may be modified by experience such as use of a tool to interact with the far space. The utility of the modeling approach relies on several aspects: (i) The two models, although devoted to different problems and simulating different brain regions, share some common mechanisms (lateral inhibition and excitation, non-linear neuron characteristics, recurrent connections, competition, Hebbian rules of potentiation and depression) that may govern more generally the fusion of senses in the brain, and the learning and plasticity of multisensory integration. (ii) The models may help interpretation of behavioral and psychophysical responses in terms of neural activity and synaptic connections. (iii) The models can make testable predictions that can help guiding future experiments in order to validate, reject, or modify the main assumptions.

Connectional parameters determine multisensory processing in a spiking network model of multisensory convergence

For the brain to synthesize information from different sensory modalities, connections from different sensory systems must converge onto individual neurons. However, despite being the definitive, first step in the multisensory process, little is known about multisensory convergence at the neuronal level. This lack of knowledge may be due to the difficulty for biological experiments to manipulate and test the connectional parameters that define convergence. Therefore, the present study used a computational network of spiking neurons to measure the influence of convergence from two separate projection areas on the responses of neurons in a convergent area. Systematic changes in the proportion of extrinsic projections, the proportion of intrinsic connections, or the amount of local inhibitory contacts affected the multisensory properties of neurons in the convergent area by influencing (1) the proportion of multisensory neurons generated, (2) the proportion of neurons that generate integrated multisensory responses, and (3) the magnitude of multisensory integration. These simulations provide insight into the connectional parameters of convergence that contribute to the generation of populations of multisensory neurons in different neural regions as well as indicate that the simple effect of multisensory convergence is sufficient to generate multisensory properties like those of biological multisensory neurons.

Conjugate Mixture Models for Clustering Multimodal Data

The problem of multimodal clustering arises whenever the data are gathered with several physically different sensors. Observations from different modalities are not necessarily aligned in the sense there there is no obvious way to associate or compare them in some common space. A solution may consist in considering multiple clustering tasks independently for each modality. The main difficulty with such an approach is to guarantee that the unimodal clusterings are mutually consistent. In this letter, we show that multimodal clustering can be addressed within a novel framework: conjugate mixture models. These models exploit the explicit transformations that are often available between an unobserved parameter space (objects) and each of the observation spaces (sensors). We formulate the problem as a likelihood maximization task and derive the associated conjugate expectation-maximization algorithm. The convergence properties of the proposed algorithm are thoroughly investigated. Several local and global optimization techniques are proposed in order to increase its convergence speed. Two initialization strategies are proposed and compared. A consistent model selection criterion is proposed. The algorithm and its variants are tested and evaluated within the task of 3D localization of several speakers using both auditory and visual data.

Neuronal responses below firing threshold for subthreshold cross-modal enhancement

Multisensory integration (such as somatosensation-vision, gustation-olfaction) could occur even between subthreshold stimuli that in isolation do not reach perceptual awareness. For example, when a somatosensory (subthreshold) stimulus is delivered within a close spatiotemporal congruency, a visual (subthreshold) stimulus evokes a visual percept. Cross-modal enhancement of visual perception is maximal when the somatosensory stimulation precedes the visual one by tens of milliseconds. This rapid modulatory response would not be consistent with a top-down mechanism acting through higher-order multimodal cortical areas, but rather a direct interaction between lower-order unimodal areas. To elucidate the neuronal mechanisms of subthreshold cross-modal enhancement, we simulated a neural network model. In the model, lower unimodal (X, Y) and higher multimodal (M) networks are reciprocally connected by bottom-up and top-down axonal projections. The lower networks are laterally connected with each other. A pair of stimuli was presented to the lower networks, whose respective intensities were too weak to induce salient neuronal activity (population response) when presented alone. Neurons of the Y network were slightly depolarized below firing threshold when a cross-modal stimulus was presented alone to the X network. This allowed the Y network to make a rapid (within tens of milliseconds) population response when presented with a subsequent congruent stimulus. The reaction speed of the Y network was accelerated, provided that the top-down projections were strengthened. We suggest that a subthreshold (nonpopulation) response to a cross-modal stimulus, acting through interaction between lower (primary unisensory) areas, may be essential for a rapid suprathreshold (population) response to a congruent stimulus that follows. Top-down influences on cross-modal enhancement may be faster than expected, accelerating reaction speed to input, in which ongoing-spontaneous subthreshold excitation of lower-order unimodal cells by higher-order multimodal cells may play an active role.

Acquired prior knowledge modulates audiovisual integration

Orienting responses to audiovisual events in the environment can benefit markedly by the integration of visual and auditory spatial information. However, logically, audiovisual integration would only be considered successful for stimuli that are spatially and temporally aligned, as these would be emitted by a single object in space-time. As humans do not have prior knowledge about whether novel auditory and visual events do indeed emanate from the same object, such information needs to be extracted from a variety of sources. For example, expectation about alignment or misalignment could modulate the strength of multisensory integration. If evidence from previous trials would repeatedly favour aligned audiovisual inputs, the internal state might also assume alignment for the next trial, and hence react to a new audiovisual event as if it were aligned. To test for such a strategy, subjects oriented a head-fixed pointer as fast as possible to a visual flash that was consistently paired, though not always spatially aligned, with a co-occurring broadband sound. We varied the probability of audiovisual alignment between experiments. Reaction times were consistently lower in blocks containing only aligned audiovisual stimuli than in blocks also containing pseudorandomly presented spatially disparate stimuli. Results demonstrate dynamic updating of the subject's prior expectation of audiovisual congruency. We discuss a model of prior probability estimation to explain the results.

Signal detection theory, uncertainty, and Poisson-like population codes

The juxtaposition of established signal detection theory models of perception and more recent claims about the encoding of uncertainty in perception is a rich source of confusion. Are the latter simply a rehash of the former? Here, we make an attempt to distinguish precisely between optimal and probabilistic computation. In optimal computation, the observer minimizes the expected cost under a posterior probability distribution. In probabilistic computation, the observer uses higher moments of the likelihood function of the stimulus on a trial-by-trial basis. Computation can be optimal without being probabilistic, and vice versa. Most signal detection theory models describe optimal computation. Behavioral data only provide evidence for a neural representation of uncertainty if they are best described by a model of probabilistic computation. We argue that single-neuron activity sometimes suffices for optimal computation, but never for probabilistic computation. A population code is needed instead. Not every population code is equally suitable, because nuisance parameters have to be marginalized out. This problem is solved by Poisson-like, but not by Gaussian variability. Finally, we build a dictionary between signal detection theory quantities and Poisson-like population quantities.

Common Bayesian models for common cognitive issues

How can an incomplete and uncertain model of the environment be used to perceive, infer, decide and act efficiently? This is the challenge that both living and artificial cognitive systems have to face. Symbolic logic is, by its nature, unable to deal with this question. The subjectivist approach to probability is an extension to logic that is designed specifically to face this challenge. In this paper, we review a number of frequently encountered cognitive issues and cast them into a common Bayesian formalism. The concepts we review are ambiguities, fusion, multimodality, conflicts, modularity, hierarchies and loops. First, each of these concepts is introduced briefly using some examples from the neuroscience, psychophysics or robotics literature. Then, the concept is formalized using a template Bayesian model. The assumptions and common features of these models, as well as their major differences, are outlined and discussed.

Optimality in mono- and multisensory map formation

In the struggle for survival in a complex and dynamic environment, nature has developed a multitude of sophisticated sensory systems. In order to exploit the information provided by these sensory systems, higher vertebrates reconstruct the spatio-temporal environment from each of the sensory systems they have at their disposal. That is, for each modality the animal computes a neuronal representation of the outside world, a monosensory neuronal map. Here we present a universal framework that allows to calculate the specific layout of the involved neuronal network by means of a general mathematical principle, viz., stochastic optimality. In order to illustrate the use of this theoretical framework, we provide a step-by-step tutorial of how to apply our model. In so doing, we present a spatial and a temporal example of optimal stimulus reconstruction which underline the advantages of our approach. That is, given a known physical signal transmission and rudimental knowledge of the detection process, our approach allows to estimate the possible performance and to predict neuronal properties of biological sensory systems. Finally, information from different sensory modalities has to be integrated so as to gain a unified perception of reality for further processing, e.g., for distinct motor commands. We briefly discuss concepts of multimodal interaction and how a multimodal space can evolve by alignment of monosensory maps.

The optimal time window of visual-auditory integration: a reaction time analysis

The spatiotemporal window of integration has become a widely accepted concept in multisensory research: crossmodal information falling within this window is highly likely to be integrated, whereas information falling outside is not. Here we further probe this concept in a reaction time context with redundant crossmodal targets. An infinitely large time window would lead to mandatory integration, a zero-width time window would rule out integration entirely. Making explicit assumptions about the arrival time difference between peripheral sensory processing times triggered by a crossmodal stimulus set, we derive a decision rule that determines an optimal window width as a function of (i) the prior odds in favor of a common multisensory source, (ii) the likelihood of arrival time differences, and (iii) the payoff for making correct or wrong decisions; moreover, we suggest a detailed experimental setup to test the theory. Our approach is in line with the well-established framework for modeling multisensory integration as (nearly) optimal decision making, but none of those studies, to our knowledge, has considered reaction time as observable variable. The theory can easily be extended to reaction times collected under the focused attention paradigm. Possible variants of the theory to account for judgments of crossmodal simultaneity are discussed. Finally, neural underpinnings of the theory in terms of oscillatory responses in primary sensory cortices are hypothesized.