Search

Brain imaging studies of English past tense inflection have found dissociations between regular and irregular verbs, but no coherent picture has emerged to explain how these dissociations arise. Here we use synthetic brain imaging on a neural network model to provide a mechanistic account of the origins of such dissociations. The model suggests that dissociations between regional activation patterns in verb inflection emerge in an adult processing system that has been shaped through experience-dependent structural brain development. Although these dissociations appear to be between regular and irregular verbs, they arise in the model from a combination of statistical properties including frequency, relationships to other verbs, and phonological complexity, without a causal role for regularity or semantics. These results are consistent with the notion that all inflections are produced in a single associative mechanism. The model generates predictions about the patterning of active brain regions for different verbs that can be tested in future imaging studies.

Brain imaging studies of English past tense inflection have found dissociations between regular and irregular verbs, but no coherent picture has emerged to explain how these dissociations arise. Here we use synthetic brain imaging on a neural network model to provide a mechanistic account of the origins of such dissociations. The model suggests that dissociations between regional activation patterns in verb inflection emerge in an adult processing system that has been shaped through experience-dependent structural brain development. Although these dissociations appear to be between regular and irregular verbs, they arise in the model from a combination of statistical properties including frequency, relationships to other verbs, and phonological complexity, without a causal role for regularity or semantics. These results are consistent with the notion that all inflections are produced in a single associative mechanism. The model generates predictions about the patterning of active brain regions for different verbs that can be tested in future imaging studies. [ABSTRACT FROM AUTHOR]

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. [ABSTRACT FROM AUTHOR]

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.

The question of whether, and if so how, learning can be transfered from previously experienced games to novel games has recently attracted the attention of the experimental game theory literature. Existing research presumes that learning operates over actions, beliefs or decision rules. This study instead uses a connectionist approach that learns a direct mapping from game payoffs to a probability distribution over own actions. Learning is operationalized as a backpropagation rule that adjusts the weights of feedforward neural networks in the direction of increasing the probability of an agent playing a myopic best response to the last game played. One advantage of this approach is that it expands the scope of the model to any possible nxn normal-form game allowing for a comprehensive model of transfer of learning. Agents are exposed to games drawn from one of seven classes of games with significantly different strategic characteristics and then forced to play games from previously unseen classes. I find significant transfer of learning, i.e., behavior that is path-dependent, or conditional on the previously seen games. Cooperation is more pronounced in new games when agents are previously exposed to games where the incentive to cooperate is stronger than the incentive to compete, i.e., when individual incentives are aligned. Prior exposure to Prisoner's dilemma, zero-sum and discoordination games led to a significant decrease in realized payoffs for all the game classes under investigation. A distinction is made between superficial and deep transfer of learning both---the former is driven by superficial payoff similarities between games, the latter by differences in the incentive structures or strategic implications of the games. I examine whether agents learn to play the Nash equilibria of games, how they select amongst multiple equilibria, and whether they transfer Nash equilibrium behavior to unseen games. Sufficient exposure to a strategically he

The extent to which meaning is involved in reading aloud has proven an area of longstanding debate, and current computational models differ on this dimension. The connectionist triangle model proposes that normal individuals rely on semantic information for correct reading of words with atypical spelling-sound relationships, but to varying degrees. This proposed individual difference would account for the varying stage of decline at which patients with semantic dementia first show the reading impairment known as surface dyslexia. Recent neuroimaging data has provided validation of this view, showing that individual differences in degree of semantic reliance during exception word reading predict the amount of activation in left anterior temporal regions associated with semantic processing. This study aimed to establish the cognitive correlates of individual differences in semantic reliance during exception word reading. Experiment 1 used a subgrouping approach with 32 participants and found larger imageability and semantic priming effects specifically for exception word reading amongst high relative to low semantic reliance readers. High semantic reliance readers also tended to read nonwords more slowly than low semantic reliance readers. A second experiment used a regression approach with 129 readers and confirmed the relationship of degree of semantic reliance both to imageability effects in exception word reading and speed of nonword reading. Further, while the performance of the higher semantic readers revealed no significant association with semantic processing tasks, there was a negative relationship with rhyme processing tasks. We therefore speculate that differences in phonological abilities may be responsible for varying degrees of semantic reliance in reading aloud. This proposal accords with the results of functional imaging showing that higher semantic reliance during exception word reading corresponds to lower activation in left pre-central gyrus, an area associated with direct spelling sound mapping and phonological processing. Our results therefore establish the nature of systematic individual differences in degree of semantic involvement amongst normal readers, and suggest directions for future neuroimaging and computational modeling research to uncover their origins. [ABSTRACT FROM AUTHOR]

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. [ABSTRACT FROM AUTHOR]

The question of whether, and if so how, learning can be transfered from previously experienced games to novel games has recently attracted the attention of the experimental game theory literature. Existing research presumes that learning operates over actions, beliefs or decision rules. This study instead uses a connectionist approach that learns a direct mapping from game payoffs to a probability distribution over own actions. Learning is operationalized as a backpropagation rule that adjusts the weights of feedforward neural networks in the direction of increasing the probability of an agent playing a myopic best response to the last game played. One advantage of this approach is that it expands the scope of the model to any possible n × n normal-form game allowing for a comprehensive model of transfer of learning. Agents are exposed to games drawn from one of seven classes of games with significantly different strategic characteristics and then forced to play games from previously unseen classes. I find significant transfer of learning, i.e., behavior that is path-dependent, or conditional on the previously seen games. Cooperation is more pronounced in new games when agents are previously exposed to games where the incentive to cooperate is stronger than the incentive to compete, i.e., when individual incentives are aligned. Prior exposure to Prisoner's dilemma, zero-sum and discoordination games led to a significant decrease in realized payoffs for all the game classes under investigation. A distinction is made between superficial and deep transfer of learning both-the former is driven by superficial payoff similarities between games, the latter by differences in the incentive structures or strategic implications of the games. I examine whether agents learn to play the Nash equilibria of games, how they select amongst multiple equilibria, and whether they transfer Nash equilibrium behavior to unseen games. Sufficient exposure to a strategically heterogeneous set of games is found to be a necessary condition for deep learning (and transfer) across game classes. Paradoxically, superficial transfer of learning is shown to lead to better outcomes than deep transfer for a wide range of game classes. The simulation results corroborate important experimental findings with human subjects, and make several novel predictions that can be tested experimentally. [ABSTRACT FROM AUTHOR]

Deep unsupervised learning in stochastic recurrent neural networks with many layers of hidden units is a recent breakthrough in neural computation research. These networks build a hierarchy of progressively more complex distributed representations of the sensory data by fitting a hierarchical generative model. In this article we discuss the theoretical foundations of this approach and we review key issues related to training, testing and analysis of deep networks for modeling language and cognitive processing. The classic letter and word perception problem of McClelland and Rumelhart (1981) is used as a tutorial example to illustrate how structured and abstract representations may emerge from deep generative learning. We argue that the focus on deep architectures and generative (rather than discriminative) learning represents a crucial step forward for the connectionist modeling enterprise, because it offers a more plausible model of cortical learning as well as way to bridge the gap between emergentist connectionist models and structured Bayesian models of cognition.

Theories of embodied cognition suppose that perception, action and cognition are tightly intertwined and share common representations and processes. Indeed, numerous empirical studies demonstrate interaction between stimulus perception, response planning and response execution. In this paper, we present an experiment and a connectionist model that show how the Simon effect, a canonical example of perception-action congruency, can be moderated by the (cognitive representation of the) task instruction. To date, no representational account of this influence exists. In the experiment, a two-dimensional Simon task was used, with critical stimuli being colored arrows pointing in one of four directions (backward, forward, left or right). Participants stood on a Wii balance board, oriented diagonally towards the screen displaying the stimuli. They were either instructed to imagine standing on a snowboard or on a pair of skis and to respond to the stimulus color by leaning towards either the left or right foot. We expected that participants in the snowboard condition would encode these movements as forward or backward, resulting in a Simon effect on this dimension. This was confirmed by the results. The left–right congruency effect was larger in the ski condition, whereas the forward–backward congruency effect appeared only in the snowboard condition. The results can be readily accounted for by HiTEC, a connectionist model that aims at capturing the interaction between perception and action at the level of representations, and the way this interaction is mediated by cognitive control. Together, the empirical work and the connectionist model contribute to a better understanding of the complex interaction between perception, cognition, and action.

Temporal discounting denotes the fact that individuals prefer smaller rewards delivered sooner over larger rewards delivered later, often to a higher extent than suggested by normative economical theories. In this article, we identify three lines of research studying this phenomenon which aim (i) to describe temporal discounting mathematically, (ii) to explain observed choice behavior psychologically, and (iii) to predict the influence of specific factors on intertemporal decisions. We then opt for an approach integrating postulated mechanisms and empirical findings from these three lines of research. Our approach focuses on the dynamical properties of decision processes and is based on computational modeling. We present a dynamic connectionist model of intertemporal choice focusing on the role of self-control and time framing as two central factors determining choice behavior. Results of our simulations indicate that the two influences interact with each other, and we present experimental data supporting this prediction. We conclude that computational modeling of the decision process dynamics can advance the integration of different strands of research in intertemporal choice.

We introduce a computational model of the negative priming (NP) effect that includes perception, memory, attention, decision making, and action. The model is designed to provide a coherent picture across competing theories of NP and to relate psychological experiments to physiological measurements. The model is formulated in terms of an abstract dynamics of activations of features, their binding into object entities or their semantic categorization as well as related memories and implied reactions. The dynamical variables interact in a connectionist network which is shown to be adaptable to a variety of experimental paradigms. We find that selective attention can be modeled by means of inhibitory processes and by a threshold dynamics. Considering the implementation it becomes obvious that the specificity of the experimental paradigm must be taken into account when predicting the nature of the NP effect.

Mastery of language can be a struggle for some children. Amongst those that succeed in achieving this feat there is variability in proficiency. Cognitive scientists remain intrigued by this variation. A now substantial body of research suggests that language acquisition is underpinned by a child's capacity for statistical learning. Moreover, a growing body of research has demonstrated that variability in statistical learning is associated with variability in language proficiency. Yet, there is a striking lack of longitudinal data. To date, there has been no comprehensive investigation of whether a capacity for statistical learning in young children is, in fact, associated with language proficiency in subsequent years. Here we review key studies that have led to the need for this longitudinal research. Advancing the language acquisition debate via longitudinal research has the potential to transform our understanding of typical development as well as disorders such as autism, specific language impairment and dyslexia.

Errors in choice tasks are preceded by gradual changes in brain activity presumably related to fluctuations in cognitive control that promote the occurrence of errors. In the present paper, we use connectionist modeling to explore the hypothesis that these fluctuations reflect (mal-)adaptive adjustments of cognitive control. We considered ERP data from a study in which the probability of conflict in an Eriksen flanker task was manipulated in sub-blocks of trials. Errors in these data were preceded by a gradual decline of N2 amplitude. After fitting a connectionist model of conflict adaptation to the data, we analyzed simulated N2 amplitude, simulated response times and stimulus history preceding errors in the model, and found that the model produced the same pattern as obtained in the empirical data. Moreover, this pattern is not found in alternative models in which cognitive control varies randomly or in an oscillating manner. Our simulations suggest that the decline of N2 amplitude preceding errors reflects an increasing adaptation of cognitive control to specific task demands, which leads to an error when these task demands change. Taken together, these results provide evidence that error-preceding brain activity can reflect adaptive adjustments rather than unsystematic fluctuations of cognitive control, and therefore, that these errors are actually a consequence of the adaptiveness of human cognition.

The recent outburst of interest in cognitive developmental robotics is fueled by the ambition to propose ecologically plausible mechanisms of how, among other things, a learning agent/robot could ground linguistic meanings in its sensorimotor behaviour. Along this stream, we propose a model that allows the simulated iCub robot to learn the meanings of actions (point, touch and push) oriented towards objects in robot's peripersonal space. In our experiments, the iCub learns to execute motor actions and comment on them. Architecturally, the model is composed of three neural-network-based modules that are trained in different ways. The first module, a two-layer perceptron, is trained by back-propagation to attend to the target position in the visual scene, given the low-level visual information and the feature-based target information. The second module, having the form of an actor-critic architecture, is the most distinguishing part of our model, and is trained by a continuous version of reinforcement learning to execute actions as sequences, based on a linguistic command. The third module, an echo-state network, is trained to provide the linguistic description of the executed actions. The trained model generalises well in case of novel action-target combinations with randomised initial arm positions. It can also promptly adapt its behavior if the action/target suddenly changes during motor execution.