Connectionist Models of Cognition and Perception II

Preface

Contents

In previous work, we developed a neurocomputational model of list memory, based on neural mechanisms, such as recurrent self-excitation and global inhibition that implement a short-term memory activation-buffer. Here, we compare this activation-buffer with a series of mathematical buffer models that originate from the 1960s, with special emphasis on presentation rate effects. We then propose an extension of the activation-buffer to address the process of selectively updating the buffer contents, which is critical for modeling working memory and complex higher-level cognition.

Following Mirman and Spivey’s investigation [12], Musca, Rousset and Ans conducted a study on the influence of the nature of the to-be-learned material on retroactive interference (RI) in humans [13]. More RI was found for unstructured than for structured material, a result opposed to that of Mirman and Spivey [12]. This paper first presents two simulations. The first, using a three-layer backpropagation hetero-associator produced a pattern of RI results that mirrored qualitatively the structure effect on RI found in humans [13]. However the level of RI was high. In the second simulation the Dual Reverberant memory Self-Refreshing neural network model (DRSR) of Ans and Rousset [1, 2] was used. As expected, the global level of RI was reduced and the structure effect on RI was still present. We further investigated the functioning of DRSR in this situation. A proactive interference (PI) was observed, and also a structure effect on PI. Furthermore, the structure effect on RI and the structure effect on PI were negatively correlated. This trade-off between structure effect on RI and structure effect on PI found in simulation points to an interesting potential phenomenon to be investigated in humans.

The search for computational principles that underlie the functionality of different cortical areas is a fundamental scientific challenge. In the case of sensory areas, one approach is to examine how the statistical properties of natural stimuli - in the case of vision, natural images and image sequences - are related to the response properties of neurons. For simple cells, located in V1, the most prominent computational theories linking neural properties and stimulus statistics are temporal coherence and independent component analysis. For these theories, the case of spatial linear cell models has been studied in a number of recent publications, but the case of spatiotemporal models has received fairly little attention. Here we examine the spatiotemporal case by applying the theories to natural image sequence data, and by analyzing the obtained results quantitatively. We compare the properties of the spatiotemporal linear cell models learned with the methods against each other, and against parameters measured from real visual systems.

There have been many proposals of how time-to-collision is computed (see Sun & Frost [1] for a review). But the results of different tasks were not conclusive for any of these models. According to new evidence of development and tuning of tasks, we propose a simple recurrent neural network [2] to account for these phenomena. Specifically we simulated ontogenic development and tuning to speed ranges through training. Results were similar to human performance: less-trained-networks responses consistently anticipate to slow objects or large objects, and this behaviour diminishes with training.

The ability to process events in their temporal and sequential context is a fundamental skill made mandatory by constant interaction with a dynamic environment. Sequence learning studies have demonstrated that subjects exhibit detailed — and often implicit — sensitivity to the sequential structure of streams of stimuli. Current connectionist models of performance in the so-called Serial Reaction Time Task (SRT), however, fail to capture the fact that sequence learning can be based not only on sensitivity to the sequential associations between successive stimuli, but also on sensitivity to the associations between successive responses, and on the predictive relationships that exist between these sequences of responses and their effects in the environment. In this paper, we offer an initial exploration of an alternative architecture for sequence learning, based on the principles of Forward Models.

In order to draw (or copy) a character, a process of linearising takes place. In this process the complete static form of the character is broken down into a temporal sequence of strokes for graphic production. According to Thomassen, Meulenbroek and Tibosh [1], individuals develop their own production rule base, which is reflected as tendencies or strategies for graphic production. Occasionally, these principles of production come into conflict resulting in a variable sequence of production for some characters. The work described in this paper uses a connectionist modeling approach to investigate the emergence of production-based behaviours in the sequential production of characters [2]. Here, the emergence of human-like production behaviours is simulated, without the need for explicitly imposed heuristics. Demonstrating that not only are connectionist networks capable of emulating the production-sequence behaviour of humans, but also that rule-like tendencies can emerge naturally upon the basis of learning.

A control architecture of action selection, inspired by the neural loops of the dorsal part of the basal ganglia -subcortical nuclei of the vertebrate’s brain- was proved to be able to solve a minimum survival task. The present paper concerns the connection of this architecture with a navigation system. This connection is inspired by recent hypotheses concerning the role of a ventral nucleus of the basal ganglia in integrating spatial, motivational and sensorimotor information. The ventral loop selects locomotion actions generated by various navigation strategies and modulated by motivations. The dorsal loop is in charge of non-spatial task selection and of coordination with the ventral loop. Implemented in a simulated robot performing the same survival task as in the previous experiment, the whole architecture improves the robot’s survival thanks to map building and path planning abilities. Furthermore, the robot is also able to occasionally overlook the information recorded in its cognitive map in order to behave opportunistically, i.e. to reach an unexpected but visible resource, instead of a memorized but remote one. These results are discussed in terms of biological and robotic contributions.

In motor learning, two main problems arise: the missing teacher signal, and the necessity to explore high-dimensional sensorimotor spaces. Several solutions have been proposed, all of them limited in some respect. In the present work, an alternative learning mechanism is developed for the example of saccade control, implemented on a stereo vision robot camera head. This approach relies on two main principles: averaging over imperfect learning examples, and learning in multiple stages. In each stage, a saccade controller is trained with a set of imperfect learning examples. Afterwards, the output of this controller serves as starting point for the creation of a new training set with better quality. By the repetition of these steps, the controllers’ performance can be incrementally improved without the need to search from scratch for the rare learning examples with very good quality.

We present a neural network model that accounts for an observed asymmetry in the categorization of cats and dogs in 3-4 month old infants. The model establishes a link between infant behaviour and mechanisms of cortical processing. Based on developmental change in the cortex the model predicts behavioural change in infants between 3 and 10 months of age.

We present a computational model of the emergence of gaze following in infant caregiver interactions. Using the model we explore the plausibility of the hypothesis that gaze following is a skill that infants acquire because they learn that monitoring their caregiver’s direction of gaze allows them to predict where interesting objects or events in their environment are. In particular, we demonstrate that a specific basic set of mechanisms is sufficient for gaze following to emerge and we show how plausible alterations of model parameters motivated by findings on developmental disorders lead to impairments in the learning of gaze following.

Autism is a complex neurodevelopmental condition causing difficulties in social interaction, communication, behavioural flexibility and superior performance in certain other abilities. Connectionist models are potentially useful tools to explore the differences in the developmental trajectories of normal and autistic children and adults. Several models are reviewed and a suggestion for a synthesis is suggested within a biologically constrained connectionist framework.

Through brain imaging studies and studies of brain-lesioned patients with face or object recognition deficits, the fusiform face area (FFA) has been identified as a face-specific processing area. Recent work, however, illustrates that the FFA is also responsive to a wide variety of non-face objects if levels of discrimination and expertise are controlled. The mystery is why an expertise area, whose initial domain of expertise is presumably faces, would be recruited for these other domains. Here we show that features tuned for fine-level discrimination within one visually homogeneous class have high-variance responses across that class. This variability generalizes to other homogenous classes, providing a foothold for learning.

This study is aimed at detecting factors influencing perceptual feature creation. By teaching several new perceptual categories, we demonstrate the emergence of new internal representations. We focus on contrasting the role of two basic factors that govern feature creation: the informative value and the degree of parsimony of the feature set. Two methods of exploring the structure of internal features are developed using an artificial neural network. These methods were empirically implemented in two experiments, both demonstrating a preference for parsimonious internal representations, even at the expense of the informative value of the feature. Our results suggest that feature parsimony is maintained not only to optimize the resource management of the perceptual system but also to aid future category learning.

Recent results from neurophysiological studies [11] suggest that energy spectra (i.e., the square of the amplitude spectrum) could be a suitable way to simulate, in a physiologically plausible manner, the spectral integration of sensory neurons. In this paper, we show for a high-level cognition task, the adequateness of the energy spectrum as an image descriptor for neural network computations. We used a simulation of cortical complex cell functions as a perceptual model which extracts image information. In a first simulation, we tested the energy spectrum descriptors with a back-propagation auto-encoder. In a second simulation, we tested the same descriptors with a standard back-propagation heteroassociator. The results show a reliable ability of these two types of neural networks to categorize and to generalize prior training to new exemplars based on the information provided by the energy spectrum of natural scene images.

Consciousness is best approached through attention. An engineering control approach to attention and motor response is presented here, culminating in the CODAM model of consciousness. This is briefly described, as is support arising from brain dynamics, especially that for the attentional blink.

We present an extension of the Selective Attention for Identification model (SAIM) [1] in which feature extraction processes are incorporated. We show that the new version successfully models experimental results from visual search. We also predict the influence of a target cue on search. This extended version of SAIM may provide a powerful framework for understanding human visual attention.

One of the most prominent experimental paradigms for investigating the deployment of attention over time is the Attentional Blink (AB). Although there is now a great deal known about it, computational modeling of the AB remains only lightly explored. This paper responds to this limitation by proposing a prototype neural network model of the blink. A central aspect of which is a realization of the concept of consolidation into working memory, which is at the heart of the majority of current explanations of the blink.

Gluck and Bower’s¹ configural cue model is a network that represents stimuli using independent nodes for each feature and feature combination within the stimulus. One of its main limitations is the lack of any clear method for incorporating secondary learning processes such as selective attention. A new configural cue model is proposed in which node activation is dependent on the average characteristics of a dimensional sampling process. This process may be described in terms of a Markov process. Learning algorithms are used to alter the matrix of transition probabilities governing the behavior of the sampling process on each trial. This allows the model to qualitatively simulate learning effects that seem to be based on limited-capacity dimensional attention. The approach used also allows the model to be used to simulate attention learning and associative learning with feature-based stimuli. This represents a potential advance over many models used in category learning research where dominant models are either only applicable to stimuli that do not vary in terms of their dimensionality (such as ALCOVE²), or make use of stimulus representations that are incapable of learning nonlinear discriminations (such as EXIT³).

The current model is an adaptation of [1], extending it to draw more complex and abstract analogies. Units are connected by two types of modifiable connections: fast connections which transmit the current activation of the units and slow connections which implement a delay transmitting an earlier activation state of the network. The fast connections drive the network into attractor states corresponding to objects. The slow connections implement transformations between states by pushing the network out of its stable state and into another attractor basin. The fast and slow connections work together to move the network from one attractor state to another in an ordered way. Since the network can learn transformations between more than two objects we suggest how the network could draw analogies involving more than two objects.

Many aspects of human and animal behaviour require individuals to learn quickly how to classify the patterns they encounter. One might imagine that evolution by natural selection would result in neural systems emerging that are very good at learning things like this. Explicit simulations of the evolution of simple developmental neural systems confirm that such rational behaviour can indeed emerge quite easily. However, the same simulations also reveal that there are situations in which evolution seems to let the species down, and populations emerge that appear to perform rather irrationally. There are actually many reasons why this might happen. I shall present the results from a selection of my simulations that begin to explore the issues involved.

This paper provides a connectionist account of the processes underlying the multiple inference model of person impression formation proposed by Reeder, Kumar, Hesson-McInnis and Trafimow [7]. First, in a replication and extension of one of their main studies, I found evidence for discounting of trait inferences when facilitating situational forces were present consistent with earlier causality-based theories, while at the same time I replicated the lack of discounting in moral inferences as documented and predicted by Reeder et al. [7]. Second, to provide an account of how these different and sometimes contradictory inferences are formed and integrated in a coherent person impression, I present a recurrent network model that automatically integrates these inferences, resulting in a pattern that closely reproduces the observed data.

The distinct computational properties of spiking neural networks are increasingly the focus of research in computational neuroscience. When modelling these networks efficiency issues are critical. In this paper we present several algorithms for the event-driven simulation of spiking neural networks on single processor systems, which facilitate the simulation of large, highly active networks.

We look at what role there might be for sublexical units in lexical representation in the modeling of isolated visual word recognition and in the reading of text. A variety of psycholinguistic paradigms have been used to investigate exactly how much information about letter order is required to recognize a word. We review some of the phenomena and some of the modeling solutions before suggesting an anatomically-based input representation that is capable of capturing important phenomena in reading in a parsimonious way.

Native English speakers include irregular plurals in English noun-noun compounds (e.g. mice chaser) more frequently than regular plurals (e.g. *rats chaser) (Gordon, 1985). This dissociation in inflectional morphology has been argued to stem from an internal and innate morphological constraint as it is thought that the input to which English speaking children are exposed is insufficient to signal that regular plurals are prohibited in compounds but irregulars might be allowed (Marcus, Brinkmann, Clahsen, Wiese & Pinker, 1995). In addition, this dissociation in English compounds has been invoked to support the idea that regular and irregular morphology are mediated by separate cognitive systems (Pinker, 1999). The evidence of the simple recurrent networks (SRNs) presented here is used to support an alternative view that the constraint on English compounds can be derived from the general frequencies and patterns in which the two types of plural (regular and irregular) in conjunction with the possessive morpheme occur in the input.

In this study we introduce the Accumulation of Expectations technique to build vectorial representations of the orthographic and phonetic forms of all the words in a language for use in connectionist models. We demonstrate how this technique can be used to build realistic orthographic representations for all Dutch and English words from the CELEX database, and realistic phonetic representations for all Dutch words in CELEX.

Some connectionist models of speech segmentation have exploited the utterance boundary strategy, where the fact that utterance endings are also word endings is used to infer where word boundaries are. In this paper, it is demonstrated that using a simple N-gram based approach outperforms the neural networks for bigrams and especially for trigrams. Moreover the trigrams performance is better than that reported for all but the best 3 unsupervised models of speech segmentation in the literature. The implications of these findings both for connectionist models of segmentation and for the cognitive modelling of segmentation more generally are discussed.

We describe the principles of designing an oscillatory neural network for processing visual information and show that the functioning of the system can be based on two main principles: the synchronization of oscillators via phase-locking and resonant increase of the amplitudes of oscillators if they work in-phase with other oscillators. The processing of a printed word is presented as an example of computer simulation.

In order to explore underlying brain mechanisms and to further understand how and where object feature binding occurs, psychophysical data are analysed and will be modelled using an attractor network. This paper describes psychophysical work and an outline of the proposed model. A rapid serial visual processing paradigm with a post-cue response task was used in three experimental conditions: spatial, temporal and spatio-temporal. Using a ‘staircase’, stimulus onset asynchrony for each observer for each condition was set in practice trails to achieve 50% error rates. Results indicate that spatial location information helps bind objects features and temporal location information hinders it. Our expectation is that the proposed neural model will demonstrate a binding mechanism by exhibiting regions of enhanced activity in the location of the target when presented with a partial post-cue. In future work, the model could be lesioned so that neuropsychological phenomena might be exhibited. In such a way, the mechanisms underlying object feature binding might be clarified.