Learning Orthographic Structure With Sequential Generative Neural Networks

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.

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.

Frequency-tagged visual evoked responses track syllable effects in visual word recognition

The processing of syllables in visual word recognition was investigated using a novel paradigm based on steady-state visual evoked potentials (SSVEPs). French words were presented to proficient readers in a delayed naming task. Words were split into two segments, the first of which was flickered at 18.75 Hz and the second at 25 Hz. The first segment either matched (congruent condition) or did not match (incongruent condition) the first syllable. The SSVEP responses in the congruent condition showed increased power compared to the responses in the incongruent condition, providing new evidence that syllables are important sublexical units in visual word recognition and reading aloud. With respect to the neural correlates of the effect, syllables elicited an early activation of a right hemisphere network. This network is typically associated with the programming of complex motor sequences, cognitive control and timing. Subsequently, responses were obtained in left hemisphere areas related to phonological processing.

Steady state visual evoked potentials in reading aloud: Effects of lexicality, frequency and orthographic familiarity

The present study explored the possibility to use Steady-State Visual Evoked Potentials (SSVEPs) as a tool to investigate the core mechanisms in visual word recognition. In particular, we investigated three benchmark effects of reading aloud: lexicality (words vs. pseudowords), frequency (high-frequency vs. low-frequency words), and orthographic familiarity ('familiar' versus 'unfamiliar' pseudowords). We found that words and pseudowords elicited robust SSVEPs. Words showed larger SSVEPs than pseudowords and high-frequency words showed larger SSVEPs than low-frequency words. SSVEPs were not sensitive to orthographic familiarity. We further localized the neural generators of the SSVEP effects. The lexicality effect was located in areas associated with early level of visual processing, i.e. in the right occipital lobe and in the right precuneus. Pseudowords produced more activation than words in left sensorimotor areas, rolandic operculum, insula, supramarginal gyrus and in the right temporal gyrus. These areas are devoted to speech processing and/or spelling-to-sound conversion. The frequency effect involved the left temporal pole and orbitofrontal cortex, areas previously implicated in semantic processing and stimulus-response associations respectively, and the right postcentral and parietal inferior gyri, possibly indicating the involvement of the right attentional network.

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.

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.