Modeling language and cognition with deep unsupervised learning: a tutorial overview

Federated inference and belief sharing

This paper concerns the distributed intelligence or federated inference that emerges under belief-sharing among agents who share a common world-and world model. Imagine, for example, several animals keeping a lookout for predators. Their collective surveillance rests upon being able to communicate their beliefs-about what they see-among themselves. But, how is this possible? Here, we show how all the necessary components arise from minimising free energy. We use numerical studies to simulate the generation, acquisition and emergence of language in synthetic agents. Specifically, we consider inference, learning and selection as minimising the variational free energy of posterior (i.e., Bayesian) beliefs about the states, parameters and structure of generative models, respectively. The common theme-that attends these optimisation processes-is the selection of actions that minimise expected free energy, leading to active inference, learning and model selection (a.k.a., structure learning). We first illustrate the role of communication in resolving uncertainty about the latent states of a partially observed world, on which agents have complementary perspectives. We then consider the acquisition of the requisite language-entailed by a likelihood mapping from an agent's beliefs to their overt expression (e.g., speech)-showing that language can be transmitted across generations by active learning. Finally, we show that language is an emergent property of free energy minimisation, when agents operate within the same econiche. We conclude with a discussion of various perspectives on these phenomena; ranging from cultural niche construction, through federated learning, to the emergence of complexity in ensembles of self-organising systems.

A Developmental Approach for Training Deep Belief Networks

Deep belief networks (DBNs) are stochastic neural networks that can extract rich internal representations of the environment from the sensory data. DBNs had a catalytic effect in triggering the deep learning revolution, demonstrating for the very first time the feasibility of unsupervised learning in networks with many layers of hidden neurons. These hierarchical architectures incorporate plausible biological and cognitive properties, making them particularly appealing as computational models of human perception and cognition. However, learning in DBNs is usually carried out in a greedy, layer-wise fashion, which does not allow to simulate the holistic maturation of cortical circuits and prevents from modeling cognitive development. Here we present iDBN, an iterative learning algorithm for DBNs that allows to jointly update the connection weights across all layers of the model. We evaluate the proposed iterative algorithm on two different sets of visual stimuli, measuring the generative capabilities of the learned model and its potential to support supervised downstream tasks. We also track network development in terms of graph theoretical properties and investigate the potential extension of iDBN to continual learning scenarios. DBNs trained using our iterative approach achieve a final performance comparable to that of the greedy counterparts, at the same time allowing to accurately analyze the gradual development of internal representations in the deep network and the progressive improvement in task performance. Our work paves the way to the use of iDBN for modeling neurocognitive development.

The secret life of predictive brains: what's spontaneous activity for?

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.

Learning Numerosity Representations with Transformers: Number Generation Tasks and Out-of-Distribution Generalization

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.

The phase space of meaning model of psychopathology: A computer simulation modelling study

INTRODUCTION

The hypothesis of a general psychopathology factor that underpins all common forms of mental disorders has been gaining momentum in contemporary clinical research and is known as the p factor hypothesis. Recently, a semiotic, embodied, and psychoanalytic conceptualisation of the p factor has been proposed called the Harmonium Model, which provides a computational account of such a construct. This research tested the core tenet of the Harmonium model, which is the idea that psychopathology can be conceptualised as due to poorly-modulable cognitive processes, and modelled the concept of Phase Space of Meaning (PSM) at the computational level.

METHOD

Two studies were performed, both based on a simulation design implementing a deep learning model, simulating a cognitive process: a classification task. The level of performance of the task was considered the simulated equivalent to the normality-psychopathology continuum, the dimensionality of the neural network's internal computational dynamics being the simulated equivalent of the PSM's dimensionality.

RESULTS

The neural networks' level of performance was shown to be associated with the characteristics of the internal computational dynamics, assumed to be the simulated equivalent of poorly-modulable cognitive processes.

DISCUSSION

Findings supported the hypothesis. They showed that the neural network's low performance was a matter of the combination of predicted characteristics of the neural networks' internal computational dynamics. Implications, limitations, and further research directions are discussed.

A Database for Learning Numbers by Visual Finger Recognition in Developmental Neuro-Robotics

Numerical cognition is a fundamental component of human intelligence that has not been fully understood yet. Indeed, it is a subject of research in many disciplines, e.g., neuroscience, education, cognitive and developmental psychology, philosophy of mathematics, linguistics. In Artificial Intelligence, aspects of numerical cognition have been modelled through neural networks to replicate and analytically study children behaviours. However, artificial models need to incorporate realistic sensory-motor information from the body to fully mimic the children's learning behaviours, e.g., the use of fingers to learn and manipulate numbers. To this end, this article presents a database of images, focused on number representation with fingers using both human and robot hands, which can constitute the base for building new realistic models of numerical cognition in humanoid robots, enabling a grounded learning approach in developmental autonomous agents. The article provides a benchmark analysis of the datasets in the database that are used to train, validate, and test five state-of-the art deep neural networks, which are compared for classification accuracy together with an analysis of the computational requirements of each network. The discussion highlights the trade-off between speed and precision in the detection, which is required for realistic applications in robotics.

Visual sense of number vs. sense of magnitude in humans and machines

Numerosity perception is thought to be foundational to mathematical learning, but its computational bases are strongly debated. Some investigators argue that humans are endowed with a specialized system supporting numerical representations; others argue that visual numerosity is estimated using continuous magnitudes, such as density or area, which usually co-vary with number. Here we reconcile these contrasting perspectives by testing deep neural networks on the same numerosity comparison task that was administered to human participants, using a stimulus space that allows the precise measurement of the contribution of non-numerical features. Our model accurately simulates the psychophysics of numerosity perception and the associated developmental changes: discrimination is driven by numerosity, but non-numerical features also have a significant impact, especially early during development. Representational similarity analysis further highlights that both numerosity and continuous magnitudes are spontaneously encoded in deep networks even when no task has to be carried out, suggesting that numerosity is a major, salient property of our visual environment.

The Challenge of Modeling the Acquisition of Mathematical Concepts

As a full-blown research topic, numerical cognition is investigated by a variety of disciplines including cognitive science, developmental and educational psychology, linguistics, anthropology and, more recently, biology and neuroscience. However, despite the great progress achieved by such a broad and diversified scientific inquiry, we are still lacking a comprehensive theory that could explain how numerical concepts are learned by the human brain. In this perspective, I argue that computer simulation should have a primary role in filling this gap because it allows identifying the finer-grained computational mechanisms underlying complex behavior and cognition. Modeling efforts will be most effective if carried out at cross-disciplinary intersections, as attested by the recent success in simulating human cognition using techniques developed in the fields of artificial intelligence and machine learning. In this respect, deep learning models have provided valuable insights into our most basic quantification abilities, showing how numerosity perception could emerge in multi-layered neural networks that learn the statistical structure of their visual environment. Nevertheless, this modeling approach has not yet scaled to more sophisticated cognitive skills that are foundational to higher-level mathematical thinking, such as those involving the use of symbolic numbers and arithmetic principles. I will discuss promising directions to push deep learning into this uncharted territory. If successful, such endeavor would allow simulating the acquisition of numerical concepts in its full complexity, guiding empirical investigation on the richest soil and possibly offering far-reaching implications for educational practice.

Numerosity discrimination in deep neural networks: Initial competence, developmental refinement and experience statistics

Both humans and non-human animals exhibit sensitivity to the approximate number of items in a visual array, as indexed by their performance in numerosity discrimination tasks, and even neonates can detect changes in numerosity. These findings are often interpreted as evidence for an innate 'number sense'. However, recent simulation work has challenged this view by showing that human-like sensitivity to numerosity can emerge in deep neural networks that build an internal model of the sensory data. This emergentist perspective posits a central role for experience in shaping our number sense and might explain why numerical acuity progressively increases over the course of development. Here we substantiate this hypothesis by introducing a progressive unsupervised deep learning algorithm, which allows us to model the development of numerical acuity through experience. We also investigate how the statistical distribution of numerical and non-numerical features in natural environments affects the emergence of numerosity representations in the computational model. Our simulations show that deep networks can exhibit numerosity sensitivity prior to any training, as well as a progressive developmental refinement that is modulated by the statistical structure of the learning environment. To validate our simulations, we offer a refinement to the quantitative characterization of the developmental patterns observed in human children. Overall, our findings suggest that it may not be necessary to assume that animals are endowed with a dedicated system for processing numerosity, since domain-general learning mechanisms can capture key characteristics others have attributed to an evolutionarily specialized number system.

Emergence of Network Motifs in Deep Neural Networks

Network science can offer fundamental insights into the structural and functional properties of complex systems. For example, it is widely known that neuronal circuits tend to organize into basic functional topological modules, called network motifs. In this article, we show that network science tools can be successfully applied also to the study of artificial neural networks operating according to self-organizing (learning) principles. In particular, we study the emergence of network motifs in multi-layer perceptrons, whose initial connectivity is defined as a stack of fully-connected, bipartite graphs. Simulations show that the final network topology is shaped by learning dynamics, but can be strongly biased by choosing appropriate weight initialization schemes. Overall, our results suggest that non-trivial initialization strategies can make learning more effective by promoting the development of useful network motifs, which are often surprisingly consistent with those observed in general transduction networks.

Hybrid IRBM-BPNN Approach for Error Parameter Estimation of SINS on Aircraft

To realize the error parameter estimation of strap-down inertial navigation system (SINS) and improve the navigation accuracy for aircraft, a hybrid improved restricted Boltzmann machine BP neural network (IRBM-BPNN) approach, which combines restricted Boltzmann machine (RBM) and BP neural network (BPNN), is proposed to forecast the inertial measurement unit (IMU) instrument errors and initial alignment errors of SINS. Firstly, the error generation mechanism of SINS is analyzed, and initial alignment error model and IMU instrument error model are established. Secondly, an unsupervised RBM method is introduced to initialize BPNN to improve the forecast performance of the neural network. The RBM-BPNN model is constructed through the information fusion of SINS/GPS/CNS integrated navigation system by using the sum of position deviation, the sum of velocity deviation and the sum of attitude deviation as the inputs and by using the error parameters of SINS as the outputs. The RBM-BPNN structure is improved to enhance its forecast accuracy, and the pulse signal is increased as the input of the neural network. Finally, we conduct simulation experiments to forecast and compensate the error parameters of the proposed IRBM-BPNN method. Simulation results show that the artificial neural network method is feasible and effective in forecasting SINS error parameters, and the forecast accuracy of SINS error parameters can be effectively improved by combining RBM and BPNN methods and improving the neural network structure. The proposed IRBM-BPNN method has the optimal forecast accuracy of SINS error parameters and navigation accuracy of aircraft compared with the radial basis function neural network method and BPNN method.

A deep learning method to more accurately recall known lysine acetylation sites

Background

Lysine acetylation in protein is one of the most important post-translational modifications (PTMs). It plays an important role in essential biological processes and is related to various diseases. To obtain a comprehensive understanding of regulatory mechanism of lysine acetylation, the key is to identify lysine acetylation sites. Previously, several shallow machine learning algorithms had been applied to predict lysine modification sites in proteins. However, shallow machine learning has some disadvantages. For instance, it is not as effective as deep learning for processing big data.

Results

In this work, a novel predictor named DeepAcet was developed to predict acetylation sites. Six encoding schemes were adopted, including a one-hot, BLOSUM62 matrix, a composition of K-space amino acid pairs, information gain, physicochemical properties, and a position specific scoring matrix to represent the modified residues. A multilayer perceptron (MLP) was utilized to construct a model to predict lysine acetylation sites in proteins with many different features. We also integrated all features and implemented the feature selection method to select a feature set that contained 2199 features. As a result, the best prediction achieved 84.95% accuracy, 83.45% specificity, 86.44% sensitivity, 0.8540 AUC, and 0.6993 MCC in a 10-fold cross-validation. For an independent test set, the prediction achieved 84.87% accuracy, 83.46% specificity, 86.28% sensitivity, 0.8407 AUC, and 0.6977 MCC.

Conclusion

The predictive performance of our DeepAcet is better than that of other existing methods. DeepAcet can be freely downloaded from https://github.com/Sunmile/DeepAcet.

Electronic supplementary material

The online version of this article (10.1186/s12859-019-2632-9) contains supplementary material, which is available to authorized users.

Children's scale errors are a natural consequence of learning to associate objects with actions: A computational model

Young children sometimes attempt an action on an object, which is inappropriate because of the object size-they make scale errors. Existing theories suggest that scale errors may result from immaturities in children's action planning system, which might be overpowered by increased complexity of object representations or developing teleofunctional bias. We used computational modelling to emulate children's learning to associate objects with actions and to select appropriate actions, given object shape and size. A computational Developmental Deep Model of Action and Naming (DDMAN) was built on the dual-route theory of action selection, in which actions on objects are selected via a direct (nonsemantic or visual) route or an indirect (semantic) route. As in case of children, DDMAN produced scale errors: the number of errors was high at the beginning of training and decreased linearly but did not disappear completely. Inspection of emerging object-action associations revealed that these were coarsely organized by shape, hence leading DDMAN to initially select actions based on shape rather than size. With experience, DDMAN gradually learned to use size in addition to shape when selecting actions. Overall, our simulations demonstrate that children's scale errors are a natural consequence of learning to associate objects with actions.

Modeling the Structure and Dynamics of Semantic Processing

The contents and structure of semantic memory have been the focus of much recent research, with major advances in the development of distributional models, which use word co-occurrence information as a window into the semantics of language. In parallel, connectionist modeling has extended our knowledge of the processes engaged in semantic activation. However, these two lines of investigation have rarely been brought together. Here, we describe a processing model based on distributional semantics in which activation spreads throughout a semantic network, as dictated by the patterns of semantic similarity between words. We show that the activation profile of the network, measured at various time points, can successfully account for response times in lexical and semantic decision tasks, as well as for subjective concreteness and imageability ratings. We also show that the dynamics of the network is predictive of performance in relational semantic tasks, such as similarity/relatedness rating. Our results indicate that bringing together distributional semantic networks and spreading of activation provides a good fit to both automatic lexical processing (as indexed by lexical and semantic decisions) as well as more deliberate processing (as indexed by ratings), above and beyond what has been reported for previous models that take into account only similarity resulting from network structure.

Deep temporal models and active inference

How do we navigate a deeply structured world? Why are you reading this sentence first - and did you actually look at the fifth word? This review offers some answers by appealing to active inference based on deep temporal models. It builds on previous formulations of active inference to simulate behavioural and electrophysiological responses under hierarchical generative models of state transitions. Inverting these models corresponds to sequential inference, such that the state at any hierarchical level entails a sequence of transitions in the level below. The deep temporal aspect of these models means that evidence is accumulated over nested time scales, enabling inferences about narratives (i.e., temporal scenes). We illustrate this behaviour with Bayesian belief updating - and neuronal process theories - to simulate the epistemic foraging seen in reading. These simulations reproduce perisaccadic delay period activity and local field potentials seen empirically. Finally, we exploit the deep structure of these models to simulate responses to local (e.g., font type) and global (e.g., semantic) violations; reproducing mismatch negativity and P300 responses respectively.

A Developmental Approach to Machine Learning?

Visual learning depends on both the algorithms and the training material. This essay considers the natural statistics of infant- and toddler-egocentric vision. These natural training sets for human visual object recognition are very different from the training data fed into machine vision systems. Rather than equal experiences with all kinds of things, toddlers experience extremely skewed distributions with many repeated occurrences of a very few things. And though highly variable when considered as a whole, individual views of things are experienced in a specific order - with slow, smooth visual changes moment-to-moment, and developmentally ordered transitions in scene content. We propose that the skewed, ordered, biased visual experiences of infants and toddlers are the training data that allow human learners to develop a way to recognize everything, both the pervasively present entities and the rarely encountered ones. The joint consideration of real-world statistics for learning by researchers of human and machine learning seems likely to bring advances in both disciplines.

Computational foundations of the visual number sense

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.

Deep temporal models and active inference

How do we navigate a deeply structured world? Why are you reading this sentence first - and did you actually look at the fifth word? This review offers some answers by appealing to active inference based on deep temporal models. It builds on previous formulations of active inference to simulate behavioural and electrophysiological responses under hierarchical generative models of state transitions. Inverting these models corresponds to sequential inference, such that the state at any hierarchical level entails a sequence of transitions in the level below. The deep temporal aspect of these models means that evidence is accumulated over nested time scales, enabling inferences about narratives (i.e., temporal scenes). We illustrate this behaviour with Bayesian belief updating - and neuronal process theories - to simulate the epistemic foraging seen in reading. These simulations reproduce perisaccadic delay period activity and local field potentials seen empirically. Finally, we exploit the deep structure of these models to simulate responses to local (e.g., font type) and global (e.g., semantic) violations; reproducing mismatch negativity and P300 responses respectively.

The Role of Architectural and Learning Constraints in Neural Network Models: A Case Study on Visual Space Coding

The recent "deep learning revolution" in artificial neural networks had strong impact and widespread deployment for engineering applications, but the use of deep learning for neurocomputational modeling has been so far limited. In this article we argue that unsupervised deep learning represents an important step forward for improving neurocomputational models of perception and cognition, because it emphasizes the role of generative learning as opposed to discriminative (supervised) learning. As a case study, we present a series of simulations investigating the emergence of neural coding of visual space for sensorimotor transformations. We compare different network architectures commonly used as building blocks for unsupervised deep learning by systematically testing the type of receptive fields and gain modulation developed by the hidden neurons. In particular, we compare Restricted Boltzmann Machines (RBMs), which are stochastic, generative networks with bidirectional connections trained using contrastive divergence, with autoencoders, which are deterministic networks trained using error backpropagation. For both learning architectures we also explore the role of sparse coding, which has been identified as a fundamental principle of neural computation. The unsupervised models are then compared with supervised, feed-forward networks that learn an explicit mapping between different spatial reference frames. Our simulations show that both architectural and learning constraints strongly influenced the emergent coding of visual space in terms of distribution of tuning functions at the level of single neurons. Unsupervised models, and particularly RBMs, were found to more closely adhere to neurophysiological data from single-cell recordings in the primate parietal cortex. These results provide new insights into how basic properties of artificial neural networks might be relevant for modeling neural information processing in biological systems.

Learning representation hierarchies by sharing visual features: a computational investigation of Persian character recognition with unsupervised deep learning

In humans, efficient recognition of written symbols is thought to rely on a hierarchical processing system, where simple features are progressively combined into more abstract, high-level representations. Here, we present a computational model of Persian character recognition based on deep belief networks, where increasingly more complex visual features emerge in a completely unsupervised manner by fitting a hierarchical generative model to the sensory data. Crucially, high-level internal representations emerging from unsupervised deep learning can be easily read out by a linear classifier, achieving state-of-the-art recognition accuracy. Furthermore, we tested the hypothesis that handwritten digits and letters share many common visual features: A generative model that captures the statistical structure of the letters distribution should therefore also support the recognition of written digits. To this aim, deep networks trained on Persian letters were used to build high-level representations of Persian digits, which were indeed read out with high accuracy. Our simulations show that complex visual features, such as those mediating the identification of Persian symbols, can emerge from unsupervised learning in multilayered neural networks and can support knowledge transfer across related domains.

An emergentist perspective on the origin of number sense

The finding that human infants and many other animal species are sensitive to numerical quantity has been widely interpreted as evidence for evolved, biologically determined numerical capacities across unrelated species, thereby supporting a 'nativist' stance on the origin of number sense. Here, we tackle this issue within the 'emergentist' perspective provided by artificial neural network models, and we build on computer simulations to discuss two different approaches to think about the innateness of number sense. The first, illustrated by artificial life simulations, shows that numerical abilities can be supported by domain-specific representations emerging from evolutionary pressure. The second assumes that numerical representations need not be genetically pre-determined but can emerge from the interplay between innate architectural constraints and domain-general learning mechanisms, instantiated in deep learning simulations. We show that deep neural networks endowed with basic visuospatial processing exhibit a remarkable performance in numerosity discrimination before any experience-dependent learning, whereas unsupervised sensory experience with visual sets leads to subsequent improvement of number acuity and reduces the influence of continuous visual cues. The emergent neuronal code for numbers in the model includes both numerosity-sensitive (summation coding) and numerosity-selective response profiles, closely mirroring those found in monkey intraparietal neurons. We conclude that a form of innatism based on architectural and learning biases is a fruitful approach to understanding the origin and development of number sense.This article is part of a discussion meeting issue 'The origins of numerical abilities'.

The Emergence of Selective Attention through Probabilistic Associations between Stimuli and Actions

In this paper we show how a multilayer neural network trained to master a context-dependent task in which the action co-varies with a certain stimulus in a first context and with a second stimulus in an alternative context exhibits selective attention, i.e. filtering out of irrelevant information. This effect is rather robust and it is observed in several variations of the experiment in which the characteristics of the network as well as of the training procedure have been varied. Our result demonstrates how the filtering out of irrelevant information can originate spontaneously as a consequence of the regularities present in context-dependent training set and therefore does not necessarily depend on specific architectural constraints. The post-evaluation of the network in an instructed-delay experimental scenario shows how the behaviour of the network is consistent with the data collected in neuropsychological studies. The analysis of the network at the end of the training process indicates how selective attention originates as a result of the effects caused by relevant and irrelevant stimuli mediated by context-dependent and context-independent bidirectional associations between stimuli and actions that are extracted by the network during the learning.

Probabilistic Models and Generative Neural Networks: Towards an Unified Framework for Modeling Normal and Impaired Neurocognitive Functions

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.

Deep Learning and Developmental Learning: Emergence of Fine-to-Coarse Conceptual Categories at Layers of Deep Belief Network

In this paper, I investigate conceptual categories derived from developmental processing in a deep neural network. The similarity matrices of deep representation at each layer of neural network are computed and compared with their raw representation. While the clusters generated by raw representation stand at the basic level of abstraction, conceptual categories obtained from deep representation shows a bottom-up transition procedure. Results demonstrate a developmental course of learning from specific to general level of abstraction through learned layers of representations in a deep belief network.

Selection-for-action emerges in neural networks trained to learn spatial associations between stimuli and actions

The objects present in our environment evoke multiple conflicting actions at every moment. Thus, a mechanism that resolves this conflict is needed in order to avoid the production of chaotic ineffective behaviours. A plausible candidate for such role is the selective attention, capable of inhibiting the neural representations of the objects irrelevant in the ongoing context and as a consequence the actions they afford. In this paper, we investigated whether a selective attention mechanism emerges spontaneously during the learning of context-dependent behaviour, whereas most neurocomputational models of selective attention and action selection imply the presence of architectural constraints. To this aim, we trained a deep neural network to learn context-dependent visual-action associations. Our main result was the spontaneous emergence of an inhibitory mechanism aimed to solve conflicts between multiple afforded actions by directly suppressing the irrelevant visual stimuli eliciting the incorrect actions for the current context. This suggests that such an inhibitory mechanism emerged as a result of the incorporation of context-independent probabilistic regularities occurring between stimuli and afforded actions.

Spatial attention in written word perception

The role of attention in visual word recognition and reading aloud is a long debated issue. Studies of both developmental and acquired reading disorders provide growing evidence that spatial attention is critically involved in word reading, in particular for the phonological decoding of unfamiliar letter strings. However, studies on healthy participants have produced contrasting results. The aim of this study was to investigate how the allocation of spatial attention may influence the perception of letter strings in skilled readers. High frequency words (HFWs), low frequency words and pseudowords were briefly and parafoveally presented either in the left or the right visual field. Attentional allocation was modulated by the presentation of a spatial cue before the target string. Accuracy in reporting the target string was modulated by the spatial cue but this effect varied with the type of string. For unfamiliar strings, processing was facilitated when attention was focused on the string location and hindered when it was diverted from the target. This finding is consistent the assumptions of the CDP+ model of reading aloud, as well as with familiarity sensitivity models that argue for a flexible use of attention according with the specific requirements of the string. Moreover, we found that processing of HFWs was facilitated by an extra-large focus of attention. The latter result is consistent with the hypothesis that a broad distribution of attention is the default mode during reading of familiar words because it might optimally engage the broad receptive fields of the highest detectors in the hierarchical system for visual word recognition.

Deep generative learning of location-invariant visual word recognition

It is widely believed that orthographic processing implies an approximate, flexible coding of letter position, as shown by relative-position and transposition priming effects in visual word recognition. These findings have inspired alternative proposals about the representation of letter position, ranging from noisy coding across the ordinal positions to relative position coding based on open bigrams. This debate can be cast within the broader problem of learning location-invariant representations of written words, that is, a coding scheme abstracting the identity and position of letters (and combinations of letters) from their eye-centered (i.e., retinal) locations. We asked whether location-invariance would emerge from deep unsupervised learning on letter strings and what type of intermediate coding would emerge in the resulting hierarchical generative model. We trained a deep network with three hidden layers on an artificial dataset of letter strings presented at five possible retinal locations. Though word-level information (i.e., word identity) was never provided to the network during training, linear decoding from the activity of the deepest hidden layer yielded near-perfect accuracy in location-invariant word recognition. Conversely, decoding from lower layers yielded a large number of transposition errors. Analyses of emergent internal representations showed that word selectivity and location invariance increased as a function of layer depth. Word-tuning and location-invariance were found at the level of single neurons, but there was no evidence for bigram coding. Finally, the distributed internal representation of words at the deepest layer showed higher similarity to the representation elicited by the two exterior letters than by other combinations of two contiguous letters, in agreement with the hypothesis that word edges have special status. These results reveal that the efficient coding of written words—which was the model's learning objective—is largely based on letter-level information.