Connectionist Models of Neurocognition and Emergent Behavior

The following sections are included:

• PREFACE

• CONTENTS

This chapter introduces the topics presented in this book. The first part highlights trends and questions that form a common thread across chapters. In the second part, an overview will be given of the topics addressed in the different sections.

The emergence of any organism's brain and behaviour cannot be explained solely by examining the organism itself. We must look beyond the organism and model the ecology as well. This is the philosophical difference between the approach that we propose and the traditional approaches that seek to model behaviour and/or neural networks. We propose treating features of the ecology as independent variables and determining which parameter values are sufficient for particular types of behaviour to arise.

Biologically-inspired perceptual modelling has so far largely proceeded by mimicking known neural features or observed behaviours. However, these approaches do not explicitly model the source of the perception: the ecology. The approach described here is novel in that it focuses on the ecology of visual agents and their adaptation to their ecology, rather than being limited to imitating aspects of a particular animal's physiology or behaviour. This approach allows us to build counterfactual ecologies in which we can identify parameter values that are sufficient for the emergence of particular features of perception. Furthermore it allows us to extract the common strategies adopted by all agents in a given class of ecologies. We demonstrate the value of this approach by describing an ecology-based model of the perception of several optical illusions.

In this paper I argue for a developmental view of some of the basic visual processes in humans. The support provided for this position comes from a variety of sources. There are results from psychological, anthropological, and neuro-physiological studies here reviewed, that show how even certain low-level visual processes depend on environmental-driven learning. In addition, a neurocom-putational model of the lower visual pathway in the cortex is provided that replicates features of two main visual processing functions: orientation selectivity in V1 (primary visual cortex) and corner selectivity in V2 (secondary visual cortex). These functions emerge in the model purely by exposure to stimuli.

Visual attention can be deployed to locations within the visual array (spatial attention), to individual features such as colour (feature-based attention), or to entire objects (object-based attention). Objects are composed of features to form a perceived 'whole'. This compositional object representation reduces the storage demands by avoiding the need to store every type of object experienced. However, this approach exposes a problem of binding these constituent features (e.g. form and colour) into objects. The problem is made explicit in the higher areas of the ventral stream as information about a feature's location is absent. For feature-based attention and search, activations flow from the inferotemporal cortex to primary visual cortex without spatial cues from the dorsal stream, therefore the neural effect is applied to all locations across the visual field.^1–4

We present a model of the ventral stream (based on the Closed Loop Attention Model (CLAM) of visual search⁵) which explains this behaviour. The model also demonstrates a mechanism of binding together colour and form features from AIT through a coincidence mechanism within primary visual cortex. The visual search simulations also demonstrate how neural activations propagate from the inferotemporal cortex to enhance activity across the primary visual cortex. As CLAM is built on top of MIIND,⁶ dynamical simulations can be generated from the model using Wilson-Cowan dynamics⁷ to simulate neural populations and add a temporal aspect to the model. The simulations are realised as 2 and 3 dimensional graphical displays which allows the flow of activations for a simulation to be visualised.

A recent study¹ showed that different attention cues (social and non-social) produce qualitatively different learning effects. The mechanisms underlying such differences, however, were unclear. Here, we present a novel computational model of audio-visual learning combining two competing processes: habituation and association. The model's parameters were trained to best reproduce each infant's individual looking behavior from trial-to-trial in training and testing. We then isolated each infant's learning function to explain the variance found in preferential looking tests. The model allowed us to rigorously examine the relationship between the infants' looking behavior and their learning mechanisms. By condition, the model revealed that 8-month-olds learned faster from the social (i.e. face) than the non-social cue (i.e., flashing squares), as evidenced by the parameters of their learning functions. In general, the 4-month-olds learned more slowly than the 8-month-olds. The parameters for attention to the cue revealed that infants at both ages who weighted the social cue highly learned quickly. With non-social cues, 8-month-olds were impaired in learning, as the cue competed for attention with the target visual event Using explicit models to link looking and learning, we can draw firm conclusions about infants' cognitive development from eye-movement behavior.

The decomposition by the human visual system of visual scenes into a range of spatial frequencies is necessary for the categorization of the objects present in the visual scene. This decomposition of spatial frequencies may be particularly important for the processing of emotions. Experiments in the field of behavioral (Schyns & Oliva, 1999) and cognitive neuroscience (Vuilleumier, Armony, Driver, & Dolan, 2003) suggest that low spatial frequencies (LSF) are better than high spatial frequencies (HSF) for the categorization of emotional facial expressions (EFE). The aim of this study was to determine whether LSF information is more useful than HSF information for the categorization of emotions. We tested this hypothesis using artificial neural networks (ANN) subject to both unsupervised and supervised learning. The results indicated better emotion categorization with LSF information, thus suggesting that the HSF signal, which is also present in the BSF signal, acts as a source of noisy information during classification tasks in artificial neural systems.

Restricted Boltzmann Machines (RBMs) appear to be a good candidate to model information processing in the cerebral cortex, since they employ a simple unsupervised learning rule that can be applied to many domains and allows the training of multiple layers of representation. In this paper, we apply the RBM learning algorithm to speech perception. We show that RBMs can be used to achieve good performance in the recognition of isolated spoken digits using a multi-layer deep belief network (consisting of a number of stacked RBMs). This performance, however, appears to depend on the fine-tuning of weights with the supervised back-propagation algorithm. To investigate how central the role of back-propagation, we compare the performance of a number of deep-belief networks using fine-tuning with the performance of the same network architectures without fine-tuning. Furthermore, since one of the main strengths of RBMs is to build up multiple layers of representation, we combine the question of fine-tuning with the question of how beneficial additional layers are for the performance of the networks. To see whether the representations that emerge on higher levels make classification easier, we also apply a simple perception classification to the different levels of the deep-belief networks when it is trained without fine-tuning.

We introduce a model of word learning in infants based on cross-modal interactions. Our model employs an architecture consisting of two Self-organizing maps (SOMs), representing the visual and auditory modalities, which are connected by Hebbian links. In contrast to previous models using a similar architecture, our model employs active Hebbian connections which propagate activation between the visual and auditory maps during learning. Our results show that categorical perception emerges from these early audio-visual interactions in both domains. We argue that the learning mechanism introduced in our model could be behind the facilitation of infants categorization through verbal labelling reported in the literature.

In this paper Hebbian based cross-situational learning is incorporated into a temporal hypermap to enable it to model the acquisition of early child language from child directed speech (CDS). This model exhibits the same level of performance as an earlier non-temporal, localist, Hebbian based cross-situational model by the same author when recalling one-word utterances from associated extralinguistic information. However, the performance of the temporal model is markedly better than the non-temporal model when required to generate appropriate one-word utterances when fed with the extralinguistic information of multi-word utterances in the training corpus. Given that cognitive processes such as child language acquisition are inherently temporal in nature, this suggests that incorporating temporal processing in cognitive models may improve the performance of these models.

Visual words are peculiar cognitive entities living under several apparently conflicting requirements. Although behavioural studies have taught us just how flexible visual word representations are in the brain, the trivial fact that we can distinguish between anagrams demands that these representations somehow carry letter order. In this chapter we present a connectionist network developed so as to test a simple hypothesis on the character of the orthographic code: that its flexibility is a by-product of location invariance. We then illustrate a more explanatory aspect of the modelling approach to gain some insights into the kind of code used by this location invariant network -how it actually keeps track of letter order. Implications for the field of visual word recognition and future research are discussed.

Computational models of reading have differed in terms of whether they propose a single route forming the mapping between orthography and phonology or whether there is a lexical/sublexical route distinction. A critical test of the architecture of the reading system is how it deals with multi-letter graphemes. Rastle and Coltheart (1998) found that the presence of digraphs in nonwords but not in words led to an increase in naming times, suggesting that nonwords were processed via a distinct sequential route to words. In contrast Pagliuca, Monaghan, and McIntosh (2008) implemented a single route model of reading and showed that under conditions of visual noise the presence of digraphs in words did have an effect on naming accuracy. In this study, we investigated whether such digraph effects could be found in both words and nonwords under conditions of visual noise. If so it would suggest that effects on words and nonwords are comparable. A single route connectionist model of reading showed greater accuracy for both words and nonwords containing digraphs. Experimental results showed participants were more accurate in recognising words if they contained digraphs. However contrary to model predictions they were less accurate in recognising nonwords containing digraphs compared to controls. We discuss the challenges faced by both theoretical perspectives in interpreting these findings and in light of a psycholinguistic grain size theory of reading.

Response congruency effects occur when responses to a target stimulus are faster and/or more accurate if the correct response to that target is the same as (what would be) the correct response to a preceding prime stimulus (relative to the situation where the correct response to the prime and target differs). Computational models of lexical access are challenged by recent findings of a response congruency effect in masked primed lexical decision. Two different types of models, the activation-based Spatial Coding Model and the probability-based Bayesian Reader are reviewed. We show that neither model accommodates the empirical data. However, replacing the homogeneous inhibition in the lexical component of the Spatial Coding model with selective inhibition enables the model to account for the response congruency effect while also providing an excellent account of other masked form priming data.

It has been argued that understanding a sentence comes down to mentally simulating the state-of-affairs described by the sentence. This paper presents a particular formalization of this idea and shows how it gives rise to measures of the amount of syntactic and semantic information conveyed by each word in a sentence. These information measures predict simulated word-processing times in a connectionist model of sentence comprehension.

A computational model of free recall is described, extending previous models by including both an activation buffer (Davelaar et al., 2005) and a distributed changing context representation (Howard and Kahana, 1999, 2002). The model was used to simulate published free-recall data, and to make new predictions for the role of context in continuous distractor free recall. A faster contextual change is predicted to cause a shift from primacy to recency, in comparison to a slower contextual change. An experiment was carried out, and results are consistent with this prediction.

The pump of thought model¹ is a theoretical cell assembly model that explains how thoughts represented by assemblies can be propagated and changed by the brain. The process of human problem solving is described by this model as the transformation of thoughts through a sequence of assemblies. Activation of a critical number of neurons within the assembly leads to the activation of the entire assembly, so that manipulations on the representation of a complex object are performed. Thoughts can be represented by verbal rules. Verbal rules are based on discrete features² and allow the representation of symbolic rules. A symbolic rule contains several if-patterns and one or more then-patterns. A pattern in the context corresponds to a set of features. The similarity between rules is defined as a function of the features they have in common. A rule can establish a new assertion by the then part (its conclusion) whenever the if part (its premise) is true. We present a straightforward transformation from the symbolic rules into a distributed representation by associative memory with practical examples (diagnostic systems). Problem solving by cell assemblies is modelled by an associative memory with feedback connections.

"Response inhibition" is argued by many authors to be a general cognitive control process or function that is invoked in situations where it is necessary to avoid producing an habitual or prepotent response. Individual differences in the efficacy of this function are consequently held to underlie individual differences in performance on tasks that are thought to rely on the function. This position is supported by empirical studies which have reported mild but reliable correlations across subjects in performance on response inhibition tasks such as Stroop colour-naming and the stop-signal task. This paper investigates the computational basis of response inhibition by exploring potential common mechanisms within existing computational models of the Stroop and stop-signal tasks. It is argued that mechanisms such as lateral inhibition, which are shared by the models and which might be thought to relate to the response inhibition construct, cannot account for the observed behavioural correlations. Instead, it is suggested that such correlations are likely to arise from a computational process of attentional bias.

The top priority in high-performance manufacturing processes is the development of a new generation of control systems to enable faster, more efficient manufacturing by means of cooperative, self-organized, self-optimized behaviour. Some natural cognitive systems display effective behaviour through perception, action, deliberation, communication, and interaction with other individuals and with the environment. An artificial cognitive control architecture is presented which is based on the shared circuits model (SCM) of sociocognitive skills proposed by Hurley.¹ The proposal consists of a five-layer architecture in which the SCM approach is used to emulate such sociocognitive skills as imitation, deliberation, and mindreading. In Hurley's approach, these capacities can be enabled by mechanisms of control, mirroring the actions of others, and simulation. A control system thus designed should be capable of responding efficiently and robustly to the problems it is set. In the present implementation, the original SCM approach was enriched and modified in the light of constructive suggestions and evaluations in the literature. With these modifications, SCM served as the foundation for the design of a networked control architecture for application to an industrial case study – a high-performance drilling process. Experiments demonstrated that the proposed artificial cognitive control system can deal with nonlinearities and uncertainties in the drilling process, providing a good transient response and good error-based performance indices.

This article links a Typology of Cognitive Abilities with computer action formats to reach later cognitive remediation. For that, it is necessary to instrument a didactic and evaluative on-line teacher in his/her perception activity of the student's cognitive processes. This instrument is on-line questionnaires which bring on items of written comprehension for a student of French as a Foreign Language.

There have been many different theoretical proposals for ways of representing word meaning in a distributed fashion. We ourselves have put forward a framework for expressing aspects of lexical semantics in terms of patterns of word co-occurrences measured in large linguistic corpora. Recent advances in the modelling of fMRI measures of brain activity have started to examine patterns of activation across the cortex rather than averaging activity across a sub-volume. Mitchell et al.¹¹ have shown that simple linear models can successfully predict fMRI data from patterns of word co-occurrence for a task where participants mentally generate properties for presented word-picture pairs. Using their MRI data, we replicate their models and extend them to use our independently optimised co-occurrence patterns to demonstrate that enriched representations of word/concept meaning produce significantly better predictions of brain activity. We also explore several aspects of the parameter space underlying the supervised learning techniques used in these models.

Developmental disorders show wide variations in severity even when, on genetic grounds, it is known that there is a common underlying cause. We use connectionist models of development combined with population modelling techniques to explore possible mechanistic causes of variations in disorder severity. Specifically, we investigate the plausibility of the hypothesis that disorder variability stems from the interaction of the common cause of the disorder with variations in neurocomputational parameters also present in the typically developing population. We base our simulations on a model of developmental regression in autism, which proposes that this phenomenon arises from over-aggressive synaptic pruning¹. We simulated a population of 1000 networks in which 641 exhibited the behavioural marker of regression in their developmental trajectories in learning a notional cognitive domain. Aside from the known single cause of the disorder (an atypical connectivity pruning parameter), we then analysed which neurocomputational parameters contributed to variation in a number of characteristics of developmental regression. These included the timing of regression onset, its severity, its behavioural specificity, and the speed and extent of subsequent recovery. Results are related to existing causal frameworks that explain the origins of developmental deficits.

Previously I outlined a scheme for understanding the usefulness of computational models.¹ This scheme was accompanied by two specific proposals. Firstly, that although models have diverse purposes, the purposes of individual modelling efforts should be made explicit. Secondly, that the best use of modelling is in establishing the correspondence between model elements and empirical objects in the form of certain 'explanatory' relationships: prediction, testing, existence proofs and proofs of sufficiency and insufficiency. The current work concerns itself with empirical tests of these two claims. I survey highly cited modelling papers and from an analysis of this corpus conclude that although a diverse range of purposes are represented, neither being accompanied by an explicit statement of purpose nor being a model of my 'explanatory' type are necessary for a modelling paper to become highly cited. Neither are these factors associated with higher rates of citation. The results are situated within a philosophy of science and it is concluded that computational modelling in the cognitive sciences does not consist of a simple Popperian prediction-and-falsification dynamic. Although there may be common principles underlying model construction, they are not captured by this scheme and it is difficult to imagine how they could be captured by any simple formula.

We look at some of the assumptions made in the computational modelling of cognition. In particular, we look at some of the problems raised by the conventional modelling goal of simplicity. We review why cognitive modellers make certain choices and we emphasise the central role of abstraction. We conclude that Occam's Razor is only half the story in using implemented computational models to explain cognitive processing, and we raise a number of questions that point the way to a materialist position in computational cognitive modelling.

Shillcock et al suggest that the preoccupation with simplicity in cognitive modeling has been detrimental to the discipline. They propose instead an approach in which the complexity of the real world should be the objective of modelers. We discuss some of the difficulties in achieving this goal from the standpoint of quantitative methods of model selection.

I discuss challenges and chances offered to cognitive psychology by the recent groundbreaking progress in humanoid robot technology. The focus is on three developments. First, the robots' humanoid appearance in combination with their cognitive capabilities facilitates intuitive interaction with users and causes strong anthropomorphism and encourages the systematic investigation of the human's

"theory of robotic mind". Second, experimental investigation of interaction by means of systematic variation of robot behavior provides new approaches to investigate human behavior, which yield methodical challenges. Third, humanoid robots increasingly face similar learning and behavioral problems as humans and their performance can give insights into the structure of such problems. Finally, it has often been argued that compared to computational models robots provide an alternative way of understanding human behavior by means of synthesis of intelligent behavior. I argue that humanoid robots can provide testbeds for hypothesized models, but – regarded as models for cognition – face similar fundamental considerations in their validity from a point of view of philosophy of science as computational models in cognitive psychology do.