Search

Brains at rest generate dynamical activity that is highly structured in space and time. We suggest that spontaneous activity, as in rest or dreaming, underlies top-down dynamics of generative models. During active tasks, generative models provide top-down predictive signals for perception, cognition, and action. When the brain is at rest and stimuli are weak or absent, top-down dynamics optimize the generative models for future interactions by maximizing the entropy of explanations and minimizing model complexity. Spontaneous fluctuations of correlated activity within and across brain regions may reflect transitions between "generic priors" of the generative model: low dimensional latent variables and connectivity patterns of the most common perceptual, motor, cognitive, and interoceptive states. Even at rest, brains are proactive and predictive.

We tested the hypothesis that part of the gap in numerical competence between fish and warm-blooded vertebrates might be related to the more efficient procedures (e.g. automated conditioning chambers) used to investigate the former and could be filled by adopting an adapted version of the Skinner box in fish. We trained guppies in a visual numerosity discrimination task, featuring two difficulty levels (3 vs. 5 and 3 vs. 4) and three conditions of congruency between numerical and non-numerical cues. Unexpectedly, guppies trained with the automated device showed a much worse performance compared to previous investigations employing more “ecological” procedures. Statistical analysis indicated that the guppies overall chose the correct stimulus more often than chance; however, their average accuracy did not exceed 60% correct responses. Learning measured as performance improvement over training was significant only for the stimuli with larger numerical difference. Additionally, the target numerosity was selected more often than chance level only for the set of stimuli in which area and number were fully congruent. Re-analysis of prior studies indicate that the gap between training with the Skinner box and with a naturalistic setting was present only for numerical discriminations, but not for colour and shape discriminations. We suggest that applying automated conditioning chambers to fish might increase cognitive load and therefore interfere with achievement of numerosity discriminations.

International audience; How do children learn to read? How do deficits in various components of the reading network affect learning outcomes? How does remediating one or several components change reading performance? In this article, we summarize what is known about learning to read and how this can be formalized in a developmentally plausible computational model of reading acquisition. The model is used to understand normal and impaired reading development (dyslexia). In particular, we show that it is possible to simulate individual learning trajectories and intervention outcomes on the basis of three component skills: orthography, phonology, and vocabulary. We therefore advocate a multifactorial computational approach to understanding reading that has practical implications for dyslexia and intervention.

One of the most rapidly advancing areas of deep learning research aims at creating models that learn to disentangle the latent factors of variation from a data distribution. However, modeling joint probability mass functions is usually prohibitive, which motivates the use of conditional models assuming that some information is given as input. In the domain of numerical cognition, deep learning architectures have successfully demonstrated that approximate numerosity representations can emerge in multi-layer networks that build latent representations of a set of images with a varying number of items. However, existing models have focused on tasks requiring to conditionally estimate numerosity information from a given image. Here, we focus on a set of much more challenging tasks, which require to conditionally generate synthetic images containing a given number of items. We show that attention-based architectures operating at the pixel level can learn to produce well-formed images approximately containing a specific number of items, even when the target numerosity was not present in the training distribution.

English serves as today’s lingua franca, a role not eased by the inconsistency of its orthography. Indeed, monolingual readers of more consistent orthographies such as Italian or German learn to read more quickly than monolingual English readers. Here, we assessed whether long-lasting bilingualism would mitigate orthography-specific differences in reading speed and whether the order in which orthographies with a different regularity are learned matters. We studied high-proficiency Italian-English and English-Italian bilinguals, with at least 20 years of intensive daily exposure to the second language and its orthography and we simulated sequential learning of the two orthographies with the CDP++ connectionist model of reading. We found that group differences in reading speed were comparatively bigger with Italian stimuli than with English stimuli. Furthermore, only Italian bilinguals took advantage of a blocked presentation of Italian stimuli compared to when stimuli from both languages were presented in mixed order, suggesting a greater ability to keep language-specific orthographic representations segregated. These findings demonstrate orthographic constraints on bilingual reading, whereby the level of consistency of the first learned orthography affects later learning and performance on a second orthography. The computer simulations were consistent with these conclusions.

Previous studies suggest that associations between numbers and space are mediated by shifts of visuospatial attention along the horizontal axis. In this study, we investigated the effect of vertical shifts of overt attention, induced by optokinetic stimulation (OKS) and monitored through eye-tracking, in two tasks requiring explicit (number comparison) or implicit (parity judgment) processing of number magnitude. Participants were exposed to black-and-white stripes (OKS) that moved vertically (upward or downward) or remained static (control condition). During the OKS, participants were asked to verbally classify auditory one-digit numbers as larger/smaller than 5 (comparison task; Exp. 1) or as odd/even (parity task; Exp. 2). OKS modulated response times in both experiments. In Exp.1, upward attentional displacement decreased the Magnitude effect (slower responses for large numbers) and increased the Distance effect (slower responses for numbers close to the reference). In Exp.2, we observed a complex interaction between parity, magnitude, and OKS, indicating that downward attentional displacement slowed down responses for large odd numbers. Moreover, eye tracking analyses revealed an influence of number processing on eye movements both in Exp. 1, with eye gaze shifting downwards during the processing of small numbers as compared to large ones; and in Exp. 2, with leftward shifts after large even numbers (6,8) and rightward shifts after large odd numbers (7,9). These results provide evidence of bidirectional links between number and space and extend them to the vertical dimension. Moreover, they document the influence of visuo-spatial attention on processing of numerical magnitude, numerical distance, and parity. Together, our findings are in line with grounded and embodied accounts of numerical cognition.

Localized damage to different brain regions can cause specific cognitive deficits. However, stroke lesions can also induce modifications in the functional connectivity of intrinsic brain networks, which could be responsible for the behavioral impairment. Though resting state networks (RSNs) are typically mapped using fMRI, it has been recently shown that they can also be detected from high-density EEG. We build on a state-of-the-art approach to extract RSNs from 64-channels EEG activity in a group of right stroke patients and to identify neural predictors of their cognitive performance. Fourteen RSNs previously found in fMRI and high-density EEG studies on healthy participants were successfully reconstructed from our patients' EEG recordings. We then correlated EEG-RSNs functional connectivity with neuropsychological scores, first considering a wide frequency band (1-80 Hz) and then specific frequency ranges in order to examine the association between each EEG rhythm and the behavioral impairment. We found that visuo-spatial and motor impairments were primarily associated with the dorsal attention network, with contribution dependent on the specific EEG band. These findings are in line with the hypothesis that there is a core system of brain networks involved in specific cognitive domains. Moreover, our results pave the way for low-cost EEG-based monitoring of intrinsic brain networks' functioning in neurological patients to complement clinical-behavioral measures. ispartof: CORTEX vol:138 pages:59-71 ispartof: location:Italy status: published

Individuals with Down syndrome (DS) present reduced basic numerical skills, which have a negative impact on everyday numeracy and mathematical learning. Here, we evaluated the efficacy of the adaptive (non-commercial) computerized game “The Number Race” in improving basic numerical skills in children with DS. The experimental group (EG; N = 30, Mage-in-months 118, range 70–149) completed a training playing with “The Number Race”, whereas children in the control group (CG; N = 31, Mage-in-months 138, range 76–207) worked with software aiming at improving their reading skills. The training lasted 10 weeks with two weekly sessions of 20–30 min each. We assessed both groups’ numerical and reading skills before and immediately after the end of the training, as well as at a 3-months follow-up. We found weak evidence for post-training groups differences in terms of overall numeracy score. However, the EG displayed substantial improvements in specific numerical skills and in mental calculation, which were maintained over time, and no improvement in reading. Conversely, the CG showed improvements in their reading skills as well as in number skills but to a lesser extent compared to the EG. Overall, “The Number Race” appears as a suitable tool to improve some aspects of numeracy in DS.

Roberts (2020, Learning & Behavior, 48[2], 191–192) discussed research claiming honeybees can do arithmetic. Some readers of this research might regard such claims as unlikely. The present authors used this example as a basis for a debate on the criterion that ought to be used for publication of results or conclusions that could be viewed as unlikely by a significant number of readers, editors, or reviewers.

Augmented Reality (AR) combines real-world and computer-generated data, allowing for a dynamic presentation of visual contents while moving in the physical surroundings. Though promising for the investigation of visuospatial processing in natural scenarios, experimental research exploiting AR in this context is scarce. In this study, we aimed at testing the behavioral consequences of multitasking when walking in different directions to close landmarks while responding to AR holograms outdoors. Participants were engaged in i) a visual single-task for discriminating augmented peripheral targets, ii) a navigation single-task consisting of a sequence of short goal-directed walking periods to close augmented landmarks and iii) a dual-task combining the latter tasks. We evaluated the cost of dual-tasking on cognitive and motor performance in comparison to the single-tasks, along with a subjective assessment of mental load. Cognitive-Motor Interference (CMI) was highlighted by performance costs in both the cognitive and motor domains under dual-task. Interestingly, a discrepancy between subjective and objective measures of mental load under dual-task was observed. We conclude that the attentional load induced by multitasking can have important consequences when navigating the dynamic real world, and thus needs to be addressed in a variety of daily-living contexts. In this perspective, AR is a suitable research tool for simulating dynamic tasks outdoors, enhancing the ecological validity of cognitive investigations without sacrificing the experimental rigor of laboratory research. Additionally, it provides insights into the possible impact on attention and behavior when using wearable mobile technologies that overlap virtual data to the physical environment.

The ability to choose the larger between two numbers reflects a mature understanding of the magnitude associated with numerical symbols. The present study explores how the knowledge of the number sequence and memory capacity (verbal and visuospatial) relate to number comparison skills while controlling for cardinal knowledge. Preschool children’s (N = 140, Mage‐in‐months = 58.9, range = 41–75) knowledge of the directional property of the counting list as well as the spatial mapping of digits on the visual line were assessed. The ability to order digits on the visual line mediated the relation between memory capacity and number comparison skills while controlling for cardinal knowledge. Beyond cardinality, the knowledge of the (spatial) order of numbers marks the understanding of the magnitude associated with numbers.

Contemporary neuroscience has embraced network science and dynamical systems to study the complex and self-organized structure of the human brain. Despite the developments in non-invasive neuroimaging techniques, a full understanding of the directed interactions in whole brain networks, referred to as effective connectivity, as well as their role in the emergent brain dynamics is still lacking. The main reason is that estimating brain connectivity requires solving a formidable large-scale inverse problem from indirect and noisy measurements. Building on the dynamic causal modelling framework, the present study offers a novel method for estimating whole-brain effective connectivity from resting-state functional magnetic resonance data. To this purpose sparse estimation methods are adapted to infer the parameters of our novel model, which is based on a linearized, region-specific haemodynamic response function. The resulting algorithm, referred to as sparse DCM, is shown to compare favorably with state-of-the art methods when tested on both synthetic and real data. We also provide a graph-theoretical analysis on the whole-brain effective connectivity estimated using data from a cohort of healthy individuals, which reveals properties such as asymmetry in the connectivity structure as well as the different roles of brain areas in favoring segregation or integration.

The mental representation of brief temporal durations, when assessed in standard laboratory conditions, is highly accurate. Here we show that adding or subtracting temporal durations systematically results in strong and opposite biases, namely over-estimation for addition and under-estimation for subtraction. The difference with respect to a baseline temporal reproduction task changed across durations in an operation-specific way and survived correcting for the effect due to operation sign alone, indexing a reliable signature of arithmetic processing on time representation. A second experiment replicated these findings with a different set of stimuli. This novel behavioral marker conceptually mirrors in the time domain the representational momentum found with motion, whereby the estimated spatial position of a visual target is displaced in the direction of motion itself. This momentum effect in temporal arithmetic suggests a striking analogy between time processing and visuospatial processing, which might index the presence of common computational principles.

International audience; Introduction Reading is a highly complex task that relies on the integration of visual, orthographic, phonological and semantic information. This complexity is clearly reflected in current computational models of reading (Coltheart et al.

Background Numeracy is an area of difficulty for children with Down syndrome (DS). It has been demonstrated that The Number Race, a non-commercial adaptive computer game designed to foster basic mathematical abilities, represents a promising instrument to enhance these skills in children with DS when delivered by an expert in a clinical setting. Aims In the present study, we assessed the efficacy of The Number Race when administered at home by properly instructed and remotely supervised parents. Methods and procedures Basic numerical skills were assessed before and after training, as well as at three-months follow-up. Performance of children with DS who worked at home with the parent (PG) was compared with that of children who received the training by an expert (EG). For both groups, the training lasted ten weeks, with two weekly sessions of 20−30 min. Outcomes and results Results show that both groups improved across various measures of numerical proficiency, including the overall score of the numeracy assessment battery, while only the EG showed an improvement in a measure of mental calculation. The improvements were maintained three months after the end of the training. Conclusions and implications These findings confirm the efficacy of The Number Race and extend it to an home-based setting, whereby parents administer the training with external supervision.

Stroke causes behavioral deficits in multiple cognitive domains and there is a growing interest in predicting patient performance from neuroimaging data using machine learning techniques. Here, we investigated a deep learning approach based on convolutional neural networks (CNNs) for predicting the severity of language disorder from 3D lesion images from magnetic resonance imaging (MRI) in a heterogeneous sample of stroke patients. CNN performance was compared to that of conventional (shallow) machine learning methods, including ridge regression (RR) on the images' principal components and support vector regression. We also devised a hybrid method based on re-using CNN's high-level features as additional input to the RR model. Predictive accuracy of the four different methods was further investigated in relation to the size of the training set and the level of redundancy across lesion images in the dataset, which was evaluated in terms of location and topological properties of the lesions. The Hybrid model achieved the best performance in most cases, thereby suggesting that the high-level features extracted by CNNs are complementary to principal component analysis features and improve the model's predictive accuracy. Moreover, our analyses indicate that both the size of training data and image redundancy are critical factors in determining the accuracy of a computational model in predicting behavioral outcome from the structural brain imaging data of stroke patients.

International audience; Learning to read is foundational for literacy development, yet many children in primary school fail to become efficient readers despite normal intelligence and schooling. This condition, referred to as developmental dyslexia, has been hypothesized to occur because of deficits in vision, attention, auditory and temporal processes, and phonology and language. Here, we used a developmentally plausible computational model of reading acquisition to investigate how the core deficits of dyslexia determined individual learning outcomes for 622 children (388 with dyslexia). We found that individual learning trajectories could be simulated on the basis of three component skills related to orthography, phonology, and vocabulary. In contrast, single-deficit models captured the means but not the distribution of reading scores, and a model with noise added to all representations could not even capture the means. These results show that heterogeneity and individual differences in dyslexia profiles can be simulated only with a personalized computational model that allows for multiple deficits.

It has been shown that “visual numerosity emerges as a statistical property of images in ‘deep networks’ that learn a hierarchical generative model of the sensory input”, through unsupervised deep learning [1]. The original deep generative model was based on stochastic neurons and, more importantly, on input (image) reconstruction. Statistical analysis highlighted a correlation between the numerosity present in the input and the population activity of some neurons in the second hidden layer of the network, whereas population activity of neurons in the first hidden layer correlated with total area (i.e., number of pixels) of the objects in the image. Here we further investigate whether numerosity information can be isolated as a disentangled factor of variation of the visual input. We train in unsupervised and semi-supervised fashion a latent-space generative model that has been shown capable of disentangling relevant semantic features in a variety of complex datasets, and we test its generative performance under different conditions. We then propose an approach to the problem based on the assumption that, in order to let numerosity emerge as disentangled factor of variation, we need to cancel out the sources of variation at graphical level.

Unilateral brain damage following stroke frequently hampers the processing of contralesional space. Whether and how it also affects the processing of stimuli appearing on the same side of the lesion is still poorly understood. Three main alternative hypotheses have been proposed, namely that ipsilesional processing is functionally (i) hyperefficient, (ii) impaired, or (iii) spared. Here, we investigated ipsilesional space awareness through a computerized paradigm that exploits a manipulation of concurrent information processing demands (i.e., multitasking). Twelve chronic right-hemisphere stroke patients with a total lack of awareness for the contralesional side of space were administered a task that required the spatial monitoring of two locations within the ipsilesional hemispace. Targets were presented immediately to the right of a central fixation point (3° eccentricity), or farther to the right toward the screen edge (17° eccentricity), or on both locations. Response to target position occurred either in isolation or while performing a concurrent visual or auditory task. Results showed that most errors occurred when two targets were simultaneously presented and patients were faced with additional task demands (in the visual or auditory modalities). In the context of concurrent visual load, ipsilesional targets presented at the rightmost location were omitted more frequently than those presented closer to fixation. This pattern qualifies ipsilesional processing in right-hemisphere stroke patients as functionally impaired, arguing against the notion of ipsilesional hyperperformance, especially when under visual load.

Multitasking is ubiquitous in everyday life. It can have a detrimental effect on several cognitive abilities including spatial processing in both brain-damaged and healthy participants. The present study investigated, in healthy adults, the electrophysiological mechanisms associated with correct detection vs. misdetection of peripheral visual target(s) while processing concurrent visual or auditory stimuli. Correct responses were coupled with increased N1 amplitude under visual (i.e., intra-modal) load but not under auditory (i.e., cross-modal) load. Under visual load, error responses were associated to opposite patterns on N1/N2 components for unilateral and bilateral stimuli. In particular, errors were marked by significantly reduced N1 and N2 amplitude for the left and right visual field, respectively, whereas higher N1 amplitude was found for errors to bilateral targets. This suggests that early negative components represent the biological marker of target awareness under visual load, whereby correct target detection is grounded on a threshold criterion. These results provide an electrophysiological correlate for the allocation of capacity-limited cognitive resources during the concurrent processing of multiple and heterogeneous visual stimuli.

The ability to map numbers onto space has been widely investigated with the number line (NL) task. Accurate (linear) placement is typically preceded by a developmental phase in which children assign more space to small numbers and compress large numbers to the right part of the line, thereby resembling a biased log-like mapping. Here, we exploited a task that separately assesses direction, order and accuracy of spatial mapping (DOS task) to investigate the origin of the biased NL mapping in a sample of first and second graders. Children with reduced ordinal knowledge in the DOS task showed a more biased NL mapping, which was formally assessed using the leading mathematical models of the NL task. However, the log-like biased mapping did not emerge in the DOS task, thereby showing a lack of generalisation across different number-space mapping tasks. Both ordinal knowledge in the DOS task and linearity in the NL task related to arithmetic fluency beyond domain-general cognitive factors. We conclude that the NL task is a meter of children’s arithmetic skills rather than an expression of the mental representation of numbers.

The ability to compare the numerical magnitude of symbolic numbers represents a milestone in the development of numerical skills. However, it remains unclear how basic numerical abilities contribute to the understanding of symbolic magnitude and whether the impact of these abilities may vary when symbolic numbers are presented as number words (e.g., "six vs. eight") vs. Arabic numbers (e.g., 6 vs. 8). In the present study on preschool children, we show that comparison of number words is related to cardinality knowledge whereas the comparison of Arabic digits is related to both cardinality knowledge and the ability to spatially map numbers. We conclude that comparison of symbolic numbers in preschool children relies on multiple numerical skills and representations, which can be differentially weighted depending on the presentation format. In particular, the spatial arrangement of digits on the number line seems to scaffold the development of a "spatial route" to understanding the exact magnitude of numerals.

Spatial representation of numbers has been repeatedly associated with the development of numerical and mathematical skills. However, few studies have explored the contribution of spatial mapping to exact number representation in young children. Here we designed a novel task that allows a detailed analysis of direction, ordinality, and accuracy of spatial mapping. Preschool children, who were classified as competent counters (cardinal principle knowers), placed triplets of sequentially presented digits on the visual line. The ability to correctly order triplets tended to decrease with the larger digits. When triplets were correctly ordered, the direction of spatial mapping was predominantly oriented from left to right and the positioning of the target digits was characterized by a pattern of underestimation with no evidence of logarithmic compression. Crucially, only ordinality was associated with performance in a digit comparison task. Our results suggest that the spatial (ordinal) arrangement of digits is a powerful source of information that young children can use to construct the representation of exact numbers. Therefore, digits may acquire numerical meaning based on their spatial order on the number line.

In this study, we investigated the effects of mastering multiple scripts in handwritten character recognition by means of computational simulations. In particular, we trained a set of deep neural networks on two different datasets of handwritten characters: the HODA dataset, which is a collection of images of handwritten Persian digits, and the MNIST dataset, which contains Latin handwritten digits. We simulated native language individuals (trained on a single dataset) as well as bilingual individuals (trained on both datasets), and compared their performance in a recognition task performed under different noisy conditions. Our results show the superior performance of bilingual networks in handwritten digit recognition in comparison to the monolingual networks, thereby suggesting that mastering multiple languages might facilitate knowledge transfer across similar domains.

We provide an emergentist perspective on the computational mechanism underlying numerosity perception, its development, and the role of inhibition, based on our deep neural network model. We argue that the influence of continuous visual properties does not challenge the notion of number sense, but reveals limit conditions for the computation that yields invariance in numerosity perception. Alternative accounts should be formalized in a computational model.

International audience; Cross-language comparisons can provide important constraints on our understanding of how people read aloud. French is an interesting case because it differs from most other writing systems in that it uses a large number of multi-letter vowel graphemes and consonants that are systematically silent (i.e., do not map to any lexical phonology; e.g., (Top). Here, we developed a French version of the Connectionist Dual Process Model of Reading Aloud (CDP++) that can handle multisyllabic stimuli (up to three syllables) and has a large-scale lexicon of more than 100,000 words. We tested the model on extant data and an additional experiment examining the reading aloud of nonwords with potentially silent letters. The results from the extant data showed that the model was able to capture a number of important psycholinguistic effects in the literature and explained between 52% and 67% of the item-specific variance in two large databases. The results of the silent-letter experiment showed that, contrary to what would be predicted on the basis of lexical database statistics, people generally pronounce ``silent'' consonants in nonwords. We show that the French CDP++ model faithfully predicted this effect because it implements a linear mapping between orthography and phonology. These findings highlight the theoretical and practical significance of using computational models to help determine the processes and representations that underlie skilled reading. (C) 2014 Elsevier Inc. All rights reserved.

A milestone in numerical development is the acquisition of counting principles which allow children to exactly determine the numerosity of a given set. Moreover, a canonical left-to-right spatial layout for representing numbers also emerges during preschool. These foundational aspects of numerical competence have been extensively studied, but there is sparse knowledge about the interplay between the acquisition of the cardinality principle and spatial mapping of numbers in early numerical development. The present study investigated how these skills concurrently develop before formal schooling. Preschool children were classified according to their performance in Give-a-Number and Number-to-position tasks. Experiment 1 revealed three qualitatively different groups: (i) children who did not master the cardinality principle and lacked any consistent spatial mapping for digits, (ii) children who mastered the cardinality principle and yet failed in spatial mapping, and (iii) children who mastered the cardinality principle and displayed consistent spatial mapping. This suggests that mastery of the cardinality principle does not entail the emergence of spatial mapping. Experiment 2 confirmed the presence of these three developmental stages and investigated their relation with a digit comparison task. Crucially, only children who displayed a consistent spatial mapping of numbers showed the ability to compare digits by numerical magnitude. A congruent (i.e., numerically ordered) positioning of numbers onto a visual line as well as the concept that moving rightwards (in Western cultures) conveys an increase in numerical magnitude mark the mastery of a spatial mapping principle. Children seem to rely on this spatial organization to achieve a full understanding of the magnitude relations between digits.

The use of written symbols is a major achievement of human cultural evolution. However, how abstract letter representations might be learned from vision is still an unsolved problem 1,2 . Here, we present a large-scale computational model of letter recognition based on deep neural networks 3,4 , which develops a hierarchy of increasingly more complex internal representations in a completely unsupervised way by fitting a probabilistic, generative model to the visual input 5,6 . In line with the hypothesis that learning written symbols partially recycles pre-existing neuronal circuits for object recognition 7 , earlier processing levels in the model exploit domain-general visual features learned from natural images, while domain-specific features emerge in upstream neurons following exposure to printed letters. We show that these high-level representations can be easily mapped to letter identities even for noise-degraded images, producing accurate simulations of a broad range of empirical findings on letter perception in human observers. Our model shows that by reusing natural visual primitives, learning written symbols only requires limited, domain-specific tuning, supporting the hypothesis that their shape has been culturally selected to match the statistical structure of natural environments 8 .

Connectionist models can be characterized within the more general framework of probabilistic graphical models, which allow to efficiently describe complex statistical distributions involving a large number of interacting variables. This integration allows building more realistic computational models of cognitive functions, which more faithfully reflect the underlying neural mechanisms at the same time providing a useful bridge to higher-level descriptions in terms of Bayesian computations. Here we discuss a powerful class of graphical models that can be implemented as stochastic, generative neural networks. These models overcome many limitations associated with classic connectionist models, for example by exploiting unsupervised learning in hierarchical architectures (deep networks) and by taking into account top-down, predictive processing supported by feedback loops. We review some recent cognitive models based on generative networks, and we point out promising research directions to investigate neuropsychological disorders within this approach. Though further efforts are required in order to fill the gap between structured Bayesian models and more realistic, biophysical models of neuronal dynamics, we argue that generative neural networks have the potential to bridge these levels of analysis, thereby improving our understanding of the neural bases of cognition and of pathologies caused by brain damage.

International audience; Learning the structure of event sequences is a ubiquitous problem in cognition and particularly in language. One possible solution is to learn a probabilistic generative model of sequences that allows making predictions about upcoming events. Though appealing from a neurobiological standpoint, this approach is typically not pursued in connectionist modeling. Here, we investigated a sequential version of the restricted Boltzmann machine (RBM), a stochastic recurrent neural network that extracts high-order structure from sensory data through unsupervised generative learning and can encode contextual information in the form of internal, distributed representations. We assessed whether this type of network can extract the orthographic structure of English monosyllables by learning a generative model of the letter sequences forming a word training corpus. We show that the network learned an accurate probabilistic model of English graphotactics, which can be used to make predictions about the letter following a given context as well as to autonomously generate high-quality pseudowords. The model was compared to an extended version of simple recurrent networks, augmented with a stochastic process that allows autonomous generation of sequences, and to non-connectionist probabilistic models (n-grams and hidden Markov models). We conclude that sequential RBMs and stochastic simple recurrent networks are promising candidates for modeling cognition in the temporal domain.

Growing evidence suggests that orienting visual attention in space can influence the processing of numerical magnitude, with leftward orienting speeding up the processing of small numbers relative to larger ones and the converse for rightward orienting. The manipulation of eye movements is a convenient way to direct visuospatial attention, but several aspects of the complex relationship between eye movements, attention orienting and number processing remain unexplored. In a previous study, we observed that inducing involuntary, reflexive eye movements by means of optokinetic stimulation affected number processing only when numerical magnitude was task relevant (i.e., during magnitude comparison, but not during parity judgment; Ranzini et al., in J Cogn Psychol 27, 459-470, (2015). Here, we investigated whether processing of task-irrelevant numerical magnitude can be modulated by voluntary eye movements, and whether the type of eye movements (smooth pursuit vs. saccades) would influence this interaction. Participants tracked with their gaze a dot while listening to a digit. The numerical task was to indicate whether the digit was odd or even through non-spatial, verbal responses. The dot could move leftward or rightward either continuously, allowing tracking by smooth pursuit eye movements, or in discrete steps across a series of adjacent locations, triggering a sequence of saccades. Both smooth pursuit and saccadic eye movements similarly affected number processing and modulated response times for large numbers as a function of direction of motion. These findings suggest that voluntary eye movements redirect attention in mental number space and highlight that eye movements should play a key factor in the investigation of number-space interactions.

Unilateral Spatial Neglect, the most dramatic manifestation of contralesional space unawareness, is a highly heterogeneous syndrome. The presence of neglect is related to core spatially lateralized deficits, but its severity is also modulated by several domain-general factors (such as alertness or sustained attention) and by task demands. We previously showed that a computer-based dual-task paradigm exploiting both lateralized and non-lateralized factors (i.e., attentional load/multitasking) better captures this complex scenario and exacerbates deficits for the contralesional space after right hemisphere damage. Here we asked whether multitasking would reveal contralesional spatial disorders in chronic left-hemisphere damaged (LHD) stroke patients, a population in which impaired spatial processing is thought to be uncommon. Ten consecutive LHD patients with no signs of right-sided neglect at standard neuropsychological testing performed a computerized spatial monitoring task with and without concurrent secondary tasks (i.e., multitasking). Severe contralesional (right) space unawareness emerged in most patients under attentional load in both the visual and auditory modalities. Multitasking affected the detection of contralesional stimuli both when presented concurrently with an ipsilesional one (i.e., extinction for bilateral targets) and when presented in isolation (i.e., left neglect for right-sided targets). No spatial bias emerged in a control group of healthy elderly participants, who performed at ceiling, as well as in a second control group composed of patients with Mild Cognitive Impairment. We conclude that the pathological spatial asymmetry in LHD patients cannot be attributed to a global reduction of cognitive resources but it is the consequence of unilateral brain damage. Clinical and theoretical implications of the load-dependent lack of awareness for contralesional hemispace following LHD are discussed.

The dynamical systems' approach to cognition (Dynamicism) promises computational models that effectively embed cognitive processing within its more natural behavioral context. Dynamical cognitive models also pose difficult, analytical challenges, which motivate the development of new analytical methodology. We start by illustrating the challenge by applying two conventional analytical methods to a well-known Dynamicist model of categorical perception. We then introduce our own analysis, which works by analogy with neural stimulation methods, and which yields some novel insights into the way the model works. We then extend and apply the method to a second Dynamicist model, which captures the key psychophysical trends that emerge when humans and animals compare two numbers. The results of the analysis-which reveals units with tuning functions that are monotonically related to the magnitudes of the numbers that the agents must compare-offer a clear contribution to the contentious debate concerning the way number information is encoded in the brain.

Adaptive computer games offer an attractive method for numeracy training in young children. However, the evidence for transfer of learning to standard measures of numerical and arithmetic skills is scarce. We carried out a randomized controlled trial on a sample of preschool children of middle socio-economic status to evaluate the effectiveness of the freeware videogame “The Number Race” (Wilson et al., 2006). Children were randomly assigned to the training group or to the control group performing an alternative computer-based activity matched for duration and setting. The groups were matched for age, gender, and IQ. Training yielded large improvements in mental calculation and spatial mapping of numbers, as well as smaller improvements in the semantic representation of numbers. Our findings complement previous studies that showed beneficial effects for disadvantaged children, thereby suggesting that “The Number Race” is a valuable tool for fostering mathematical learning in the general population of young children.

The most influential theory of learning to read is based on the idea that children rely on phonological decoding skills to learn novel words. According to the self-teaching hypothesis, each successful decoding encounter with an unfamiliar word provides an opportunity to acquire word-specific orthographic information that is the foundation of skilled word recognition. Therefore, phonological decoding acts as a self-teaching mechanism or ‘built-in teacher’. However, all previous connectionist models have learned the task of reading aloud through exposure to a very large corpus of spelling–sound pairs, where an ‘external’ teacher supplies the pronunciation of all words that should be learnt. Such a supervised training regimen is highly implausible. Here, we implement and test the developmentally plausible phonological decoding self-teaching hypothesis in the context of the connectionist dual process model. In a series of simulations, we provide a proof of concept that this mechanism works. The model was able to acquire word-specific orthographic representations for more than 25 000 words even though it started with only a small number of grapheme–phoneme correspondences. We then show how visual and phoneme deficits that are present at the outset of reading development can cause dyslexia in the course of reading development.

International audience; Most models of reading aloud have been constructed to explain data in relatively complex orthographies like English and French. Here, we created an Italian version of the Connectionist Dual Process Model of Reading Aloud (CDP++) to examine the extent to which the model could predict data in a language which has relatively simple orthography-phonology relationships but is relatively complex at a suprasegmental (word stress) level. We show that the model exhibits good quantitative performance and accounts for key phenomena observed in naming studies, including some apparently contradictory findings. These effects include stress regularity and stress consistency, both of which have been especially important in studies of word recognition and reading aloud in Italian. Overall, the results of the model compare favourably to an alternative connectionist model that can learn non-linear spelling-to-sound mappings. This suggests that CDP++ is currently the leading computational model of reading aloud in Italian, and that its simple linear learning mechanism adequately captures the statistical regularities of the spelling-to-sound mapping both at the segmental and supra-segmental levels.

Individuals with Down syndrome (DS) exhibit various math difficulties which can be ascribed both to global intelligence level and/or to their atypical cognitive profile. In this light, it is crucial to investigate whether DS display deficits in basic numerical skills. In the present study, individuals with DS and two groups of typically developing (TD) children matched for mental and chronological age completed two delayed match-to-sample tasks in order to evaluate the functioning of visual enumeration skills. Children with DS showed a specific deficit in the discrimination of small numerosities (within the subitizing range) with respect to both mental and chronological age matched TD children. In contrast, the discrimination of larger numerosities, though lower than that of chronological age matched controls, was comparable to that of mental age matched controls. Finally, counting was less fluent but the understanding of cardinality seemed to be preserved in DS. These results suggest a deficit of the object tracking system underlying the parallel individuation of small numerosities and a typical – but developmentally delayed – acuity of the approximate number system for discrimination of larger numerosities.

It is often assumed that graphemes are a crucial level of orthographic representation above letters. Current connectionist models of reading, however, do not address how the mapping from letters to graphemes is learned. One major challenge for computational modeling is therefore developing a model that learns this mapping and can assign the graphemes to linguistically meaningful categories such as the onset, vowel, and coda of a syllable. Here, we present a model that learns to do this in English for strings of any letter length and any number of syllables. The model is evaluated on error rates and further validated on the results of a behavioral experiment designed to examine ambiguities in the processing of graphemes. The results show that the model (a) chooses graphemes from letter strings with a high level of accuracy, even when trained on only a small portion of the English lexicon; (b) chooses a similar set of graphemes as people do in situations where different graphemes can potentially be selected; (c) predicts orthographic effects on segmentation which are found in human data; and (d) can be readily integrated into a full-blown model of multi-syllabic reading aloud such as CDP++ (Perry, Ziegler, & Zorzi, 2010). Altogether, these results suggest that the model provides a plausible hypothesis for the kind of computations that underlie the use of graphemes in skilled reading.

It is widely believed that numbers are spatially represented from left to right on the mental number line. Whether this spatial format of representation is specific to numbers or is shared by non-numerical ordered sequences remains controversial. When healthy participants are asked to randomly generate digits they show a systematic small-number bias that has been interpreted in terms of “pseudoneglect in number space”. Here we used a random generation task to compare numerical and non-numerical order. Participants performed the task at three different pacing rates and with three types of stimuli (numbers, letters, and months). In addition to a small-number bias for numbers, we observed a bias towards “early” items for letters and no bias for months. The spatial biases for numbers and letters were rate independent and similar in size, but they did not correlate across participants. Moreover, letter generation was qualified by a systematic forward direction along the sequence, suggesting that the ordinal dimension was more salient for letters than for numbers in a task that did not require its explicit processing. The dissociation between numerical and non-numerical orders is consistent with electrophysiological and neuroimaging studies and suggests that they rely on at least partially different mechanisms.

Functional near-infrared spectroscopy (fNIRS) uses near-infrared light to measure cortical concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) held to be correlated with cognitive activity. Providing a parametric depiction of such changes in the classic form of stimulus-evoked hemodynamic responses (HRs) can be attained with this technique only by solving two problems. One problem concerns the separation of informative optical signal from structurally analogous noise generated by a variety of spurious sources, such as heart beat, respiration, and vasomotor waves. Another problem pertains to the inherent variability of HRs, which is notoriously contingent on the type of experiment, brain region monitored, and human phenotype. A novel method was devised in the present context to solve both problems based on a two-step algorithm combining the treatment of noise-only data extrapolated from a reference-channel and a Bayesian filter applied on a per-trial basis. The present method was compared to two current methods based on conventional averaging, namely, a typical averaging method and an averaging method implementing the use of a reference-channel. The result of the comparison, carried out both on artificial and real data, revealed a sensitive accuracy improvement in HR estimation using the present method relative to each of the other methods.

In their letter, Skottun and Skoyles (1) question our conclusion (2) that extra-wide letter spacing is more beneficial to dyslexic children than to normally developing children. Here, we show the pitfalls of their reasoning.

Growing evidence indicates that planning eye movements and orienting visuospatial attention share overlapping brain mechanisms. A tight link between endogenous attention and eye movements is maintained by the premotor theory, in contrast to other accounts that postulate the existence of specific attention mechanisms that modulate the activity of information processing systems. The strong assumption of equivalence between attention and eye movements, however, is challenged by demonstrations that human observers are able to keep attention on a specific location while moving the eyes elsewhere. Here we investigate whether a recurrent model of saccadic planning can account for attentional effects without requiring additional or specific mechanisms separate from the circuits that perform sensorimotor transformations for eye movements. The model builds on the basis function approach and includes a circuit that performs spatial remapping using an “internal forward model” of how visual inputs are modified as a result of saccadic movements. Simulations show that the latter circuit is crucial to account for dissociations between attention and eye movements that may be invoked to disprove the premotor theory. The model provides new insights into how spatial remapping may be implemented in parietal cortex and offers a computational framework for recent proposals that link visual stability with remapping of attention pointers.

Although the causes of dyslexia are still debated, all researchers agree that the main challenge is to find ways that allow a child with dyslexia to read more words in less time, because reading more is undisputedly the most efficient intervention for dyslexia. Sophisticated training programs exist, but they typically target the component skills of reading, such as phonological awareness. After the component skills have improved, the main challenge remains (that is, reading deficits must be treated by reading more—a vicious circle for a dyslexic child). Here, we show that a simple manipulation of letter spacing substantially improved text reading performance on the fly (without any training) in a large, unselected sample of Italian and French dyslexic children. Extra-large letter spacing helps reading, because dyslexics are abnormally affected by crowding, a perceptual phenomenon with detrimental effects on letter recognition that is modulated by the spacing between letters. Extra-large letter spacing may help to break the vicious circle by rendering the reading material more easily accessible.

Growing experimental evidence suggests that temporal events are represented on a mental time line, spatially oriented from left to right. Support for the spatial representation of time comes mostly from studies that have used spatially organized responses. Moreover, many of these studies did not avoid possible confounds attributable to target stimuli that simultaneously convey both spatial and temporal dimensions. Here we show that task-irrelevant, lateralized visuospatial primes affect auditory duration judgments. Responses to short durations were faster when the auditory target was paired with left- than with right-sided primes, whereas responses to long durations were faster when paired with right- than with left-sided primes. Thus, when the representations of physical space and time are concurrently activated, physical space may influence time even when a lateralized, spatially encoded response is not required by the task. The time-space interaction reported here cannot be ascribed to any Spatial-Temporal Association of Response Codes effect. It supports the hypothesis that the representation of time is spatially organized, with short durations represented on the left space and longer ones on the right.

The representation of numerical and non-numerical ordered sequences was investigated in children from preschool to grade 3. The child's conception of how sequence items map onto a spatial scale was tested using the Number-to-Position task (Siegler & Opfer, 2003) and new variants of the task designed to probe the representation of the alphabet (i.e., letter sequence) and the calendar year (i.e., month sequence). The representation of non-numerical order showed the same developmental pattern previously observed for numerical representation, with a logarithmic mapping in the youngest children and a shift to linear mapping in older children. Although the individual ability to position non-numerical items was related to the child's knowledge of the sequence, a significant amount of unique variance was explained by her type of number-line representation. These results suggest that the child's conception of numerical order is generalized to non-numerical sequences and that the concept of linearity is acquired in the numerical domain first and progressively extended to all ordinal sequences.

Time and space are tightly linked in the physical word. Recently, several lines of evidence have suggested that the mental representation of time might be spatial in nature. For instance, time-space interactions have been described as a strong preference to associate the past with the left space and the future with the right space. Here we review the growing evidence of interactions between time and space processing, systematized according to the type of interaction being investigated. We present the empirical findings supporting the possibility that humans represent the subjective time flow on a spatially oriented "mental time line" that is accessed through spatial attention mechanisms. The heterogeneous time-space interactions are then compared with the number-space interactions described in the numerical cognition literature. An alternative hypothesis, which maintains a common system for magnitude processing, including time, space, and number, is also discussed. Finally, we extend the discussion to the more general issue of how the representation of these concepts might be grounded into the cortical circuits that support spatial attention and sensorimotor transformations.

Numerosity estimation is phylogenetically ancient and foundational to human mathematical learning, but its computational bases remain controversial. Here we show that visual numerosity emerges as a statistical property of images in 'deep networks' that learn a hierarchical generative model of the sensory input. Emergent numerosity detectors had response profiles resembling those of monkey parietal neurons and supported numerosity estimation with the same behavioral signature shown by humans and animals.

Neuroimaging studies attempting to isolate the neural substrate of visual short-term memory in humans have concentrated on the behavior of neurons populating the posterior part of the parietal cortex as a possible source of visual short-term memory capacity limits. Using a standard change-detection task, fMRI studies have shown that maintenance of bilaterally encoded objects elicited bilateral increases of hemodynamic activation in the intra-parietal and intra-occipital sulci (IPS-IOS) proportional to the number of objects retained in visual short-term memory. We used a spatially cued variant of the change-detection task to record hemodynamic responses to unilaterally encoded objects using functional near-infrared spectroscopy (fNIRS). Electrophysiological studies that employed this task have shown that maintenance of unilaterally encoded objects elicited posterior unilateral (contralateral) increase in event-related negativity proportional to the number of objects retained in visual short-term memory. We therefore examined whether contralateral increases in oxy-hemoglobin concentration correlated with the number of retained objects. Contrary to the idea that bilateral increases in BOLD responses and unilateral increases in event-related negativity may be different reflections of the same underlying neural/functional processing, memory-related increases in oxy-hemoglobin concentration were found bilaterally even when objects had to be encoded unilaterally. The present findings suggest that EEG and fMRI/fNIRS techniques reveal distinct neural signatures of the mechanisms supporting visual short-term memory.

International audience; Most words in English have more than one syllable, yet the most influential computational models of reading aloud are restricted to processing monosyllabic words. Here, we present CDP++, a new version of the Connectionist Dual Process model (Perry, Ziegler, & Zorzi, 2007). CDP++ is able to simulate the reading aloud of mono- and disyllabic words and nonwords, and learns to assign stress in exactly the same way as it learns to associate graphemes with phonemes. CDP++ is able to simulate the monosyllabic benchmark effects its predecessor could, and therefore shows full backwards compatibility. CDP++ also accounts for a number of novel effects specific to disyllabic words, including the effects of stress regularity and syllable number. In terms of database performance, CDP++ accounts for over 49% of the reaction time variance on items selected from the English Lexicon Project, a very large database of several thousand of words. With its lexicon of over 32,000 words, CDP++ is therefore a notable example of the successful scaling-up of a connectionist model to a size that more realistically approximates the human lexical system. (C) 2010 Elsevier Inc. All rights reserved.

Neuroimaging studies of numerical cognition have pointed to the horizontal segment of the intraparietal sulcus (hIPS) as the neural correlate of numerical representations in humans. However, the specificity of hIPS for numbers remains controversial. For example, its activation during numerical comparison cannot be distinguished from activation during ordinal judgments on non-numerical sequences such as letters (Fias et al., 2007, J. Neuroscience). Based on the hypothesis that the fine-grained distinction between representations of numerical vs. letter order in hIPS might simply be invisible to conventional fMRI data analysis, we used support vector machines (SVM) to reanalyse the data of Fias et al. (2007). We show that classifiers trained on hIPS voxels can discriminate between number comparison and letter comparison, even though the two tasks produce the same metric of behaviour. Voxels discriminating between the two conditions were consistent across subjects and contribution analysis revealed maps of distinct sets of voxels implicated in the processing of numerical vs. alphabetical order in bilateral hIPS. These results reconcile the neuroimaging data with the neuropsychological evidence suggesting dissociations between numbers and other non-numerical ordered sequences, and demonstrate that multivariate analyses are fundamental to address fine-grained theoretical issues with fMRI studies.