Connectionist Models of Behaviour and Cognition II

The following sections are included:

• PREFACE

• CONTENTS

Understanding the communicative intention of a speaker is the ultimate goal of language comprehension. Yet, there is very little computational work on this topic. In this chapter a general cognitive plausible model of how an addressee can understand communicative intentions is presented in mathematical detail. The key mechanism of the model is simulated role-reversal of the addressee with the speaker, i.e., the addressee puts himself in the state of the speaker and — using his own experience about plausible intentions — computes the most likely intention in the given context. To show the model's computational effectiveness, it was implemented in a multi-agent system. In this system agents learn about which states of the world are desirable using a neural network trained with reinforcement learning. The power of simulated role-reversal in understanding communicative intention was demonstrated by depriving the utterances of speakers of all content. Employing the outlined model, the agents nevertheless accomplished a remarkable understanding of intentions using context information alone.

Behavioural and brain imaging evidence has shown that seeing objects automatically evokes "affordances", for example it tends to activate internal representations related to the execution of precision or power grips. In line with this evidence, Tucker and Ellis [1] found a compatibility effect between object size (small and large) and the kind of grip (precision and power) used to respond whether seen objects were artefacts or natural objects. This work presents a neural-network model that suggests an interpretation of these experiments in agreement with a recent theory on the general functions of prefrontal cortex. Prefrontal cortex is seen as a source of top-down bias in the competition for behavioural expression of multiple neural pathways carrying different information. The model successfully reproduces the experimental results on compatibility effects and shows how, although such a bias allows organisms to perform actions which differ from those suggested by objects' affordances, these still exert their influence on behaviour as reflected by longer reaction times.

In this paper, we present a neural architecture aimed to reproduce the qualitative properties of the mirror neurons system which encodes neural representations of actions either performed or observed. Several biological researches have emphasized some of its important aspects, for instance, the tight coupling between perception and action, the crucial role of timing (temporal information for encoding and detection), or its particular neural connectivity. We attempt to model these in a network of spiking neurons to learn the accurate temporal relationships between sensorimotor maps for action representation. After the learning, the neural connectivity efficiently induces functional capabilities in the whole network exhibiting statistics similar with observed evidences in the mirror neurons system and comparable to those of small-world networks (e.g., scale-free dynamics and hierarchical organization).

Classically, visual attention is assumed to be influenced by visual properties of objects, e.g. as assessed in visual search tasks. However, recent experimental evidence suggests that visual attention is also guided by action-related properties of objects ("affordances"),^1,2 e.g. the handle of a cup affords grasping the cup; therefore attention is drawn towards the handle. In a first step towards modelling this interaction between attention and action, we implemented the Selective Attention for Action model (SAAM). The design of SAAM is based on the Selective Attention for Identification model (SAIM).³ For instance, we also followed a soft-constraint satisfaction approach in a connectionist framework. However, SAAM's selection process is guided by locations within objects suitable for grasping them whereas SAIM selects objects based on their visual properties. In order to implement SAAM's selection mechanism two sets of constraints were implemented. The first set of constraints took into account the anatomy of the hand, e.g. maximal possible distances between fingers. The second set of constraints (geometrical constraints) considered suitable contact points on objects by using simple edge detectors. We demonstrate here that SAAM can successfully mimic human behaviour by comparing simulated contact points with experimental data.

Auto-associative networks have proven extremely useful when modelling the hypothesised function of the hippocampus in both episodic and spatial memory. To date, the majority of these models have made use of rate coded neural implementations and Hebbian plasticity rules mediated by correlations between these firing rates. However, recent neurobiological evidence suggests that synaptic plasticity in the hippocampus, and many other cortical regions, depends explicitly on the temporal relationship between afferent action potentials and efferent spiking – a formulation known as spike-timing dependent plasticity (STDP). Few attempts have been made to reconcile the STDP rule with previous models of rate coded plasticity or auto-associative network function. Further complications arise from the fact that there are many computational interpretations of the empirical data regarding STDP which can each precipitate distinct network dynamics. This paper examines an STDP implementation that has been identified by previous research as able to replicate rate coded Hebbian plasticity within a spiking, recurrent neural network. We consequently demonstrate that the STDP rule and spiking neural dynamics can be reconciled with auto-associative network function, allowing the successes of previous hippocampal models to be replicated while providing them with a firmer basis in modern neurobiology.

A model is described in which the hippocampal system receives inputs from cortical columns throughout the neocortex, uses these inputs to determine the most appropriate columns for declarative information recording in response to the current sensory experience, and generates outputs that drive that recording in the selected columns. Evidence in support of the model is described, including physiological connectivity, neuron structures and algorithms, and psychological deficits resulting from damage. Preliminary results from an electronic implementation of the model using leaky integrator neuron models learning via the LTP mechanism are described.

This simulation study explores how structural processes and synaptic consolidation during hippocampal memory replay can improve the performance of neocortical neural networks by emulating high effective connectivity in networks that have only low anatomical connectivity. We model ongoing structural plasticity such that, in each time step, a certain fraction of the unconsolidated synapses are eliminated and replaced by new synapses generated at random locations. Simultaneous replay of novel memories consolidates some of the cortical synapses according to Hebbian learning. By this procedure sparsely connected networks can become functionally equivalent to densely connected networks, thereby storing a large amount of information with a tiny number of synapses. In particular, it is possible to store up to ℭ^S ≤ log₂ n bits of information per synapse in simple networks of n neurons. This is much more than the well-known bound ℭ ≤ 0.72 bits per synapse for static networks. It turns out that sufficiently fast learning requires a significant number of silent unconsolidated synapses. Thus, with lifetime and stored memories, the number of unconsolidated synapses and thus the ability to learn will decrease gradually. This leads to the discussion of various memory-related effects such as catastrophic forgetting and Ribot gradients in retrograde amnesia.

We present a computational model for contextual deficits that underlie thought disorder in Schizophrenia. Predictions were obtained using a neurocomputational model in which lexical-semantic information is maintained in an activation buffer that can bias the interpretation of ambiguous information. Deficits in the capacity to maintain information in the buffer result in deficit in context maintenance. The predictions are shown to match results from an experimental study on Schizophrenia patients and matched controls.

A sparsely connected associative memory model is built with small-world connectivity, and trained on both random, and real-world image sets. It is found that pattern recall using real-world images can vary significantly from that of random images, and that the relationship between network wiring strategy and performance changes dramatically when training sets consist of certain types of real-world image.

Hubel and Wiesel's discoveries inspired several hierarchical models for pattern and image categorization. During the hierarchical categorization the neural network gradually reduces the information from the input layer through the output layer. The units of the output layer represent the categories. Local features are integrated into more global features in sequential transformations. In our proposed model the hierarchical neural network performs beside the categorization a similarity-based image or pattern retrieval by a backward operation. During similarity-based image retrieval, the search starts from the images represented by global features. In this representation, the set of all possible similar images is determined. In the next stage, additional information corresponding to the representation of more local feature is used to reduce this set represented by some triggered units. This procedure is repeated until the similar images can be determined.

Manipulations of encoding strength and stimulus class can lead to a simultaneous increase in hits and decrease in false alarms for a given condition in a yes/no recognition memory test. Based on signal detection theory, the strength-based 'mirror effect' is thought to involve a shift in response criterion/threshold (Type I), whereas the stimulus class effect derives from a specific ordering of the memory strength signals for presented items (Type II). We implemented both suggested mechanisms in a simple, competitive feed-forward neural network model with a learning rule related to Bayesian inference. In a single-process approach to recognition, the underlying decision axis as well as the response criteria/thresholds were derived from network activation. Initial results replicated findings in the literature and are a first step towards a more neurally explicit model of mirror effects in recognition memory tests.

Recognizing expressions is a key part of human social interaction; processing of facial expression information is largely automatic in humans, but it is a non-trivial task for a computational system. The purpose of this work is to develop computational models capable of differentiating between a range of human facial expressions. Here we use two sets of images, namely: Angry and Neutral. Raw face images are examples of high dimensional data, so here we use some dimensionality reduction techniques: Principal Component Analysis and Curvilinear Component Analysis. We preprocess the images with a bank of Gabor filters, so that important features in the face images are identified. Subsequently the faces are classified using a Support Vector Machine. We also find the effect size of the pixels for the Angry and Neutral faces. We show that it is possible to differentiate faces with a neutral expression from those with an angry expression with high accuracy. Moreover we can achieve this with data that has been massively reduced in size: in the best case the original images are reduced to just 6 dimensions.

We present a neural network model of the transition from early, perceptually-based category formation to conceptual categorization based on the learning of category labels. The model investigates the interactions between category structure and labelling success in terms of category compactness, between-category similarity, frequency of labelling and prelinguistic object knowledge. By providing accounts both of the effect of category structure on word learning and of word learning on category structure the model presents the first step of a unified account of category and word learning.

Combining experiments and modeling, we study how the discrimination of time intervals depends both on the interval duration and on contextual stimuli. Participants had to judge the temporal regularity of a sequence of standard intervals that contained a deviant interval. We find that the performance to detect the deviant increases with the number of standards preceeding the deviant and decreases with the duration of the standard. While the effect of the standard duration can be explained by an neural network model that realizes the concept of multiple synfire chains, the position effect is incorporated into the model by an in-situ averaging process. Furthermore, experiments are discussed that are critical for the predictions of the model.

In this paper we describe how information theoretic methods can be used in the context of time-varying subjective probability models in order to quantify such notions as uncertainty and surprisingness as experienced by a hypothetical observer exposed to sequences of symbolic stimuli, in particular, musical patterns. Novel measures of predictive information and predictive information rate, introduced in previous work¹ as a potential model for 'interestingness' or formal aesthetic value, are extended to account for adaptation in the observer's probability model on repeated exposure to a pattern. The system is applied to minimalist music and compared with results from previous rule-based systems of music analysis.

In this article, we consider contemporary theories of concepts, and Bayesian and self-organizing models of concept formation. After introducing the different models, we present our own experiment. It utilizes a multi-agent simulation framework, in which the emergence of a common vocabulary can be studied. In the experiment, we use jointly the self-organizing maps and probabilistic modeling of concept naming. The results of the experiments show that a common vocabulary to denote prototypical colors emerges in the agent population.

We present empirical results and an implemented computational model grounded in the Dynamic Field Theory [1] that directly addresses the second-to-second dynamics governing the integration of linguistic and non-linguistic spatial systems. Results from two experiments show activating a spatial term can differentially bias location memories in the direction of the spatial term prototype. Subsequent simulations from a hybrid Dynamic Field Theory-connectionist model capture the observed term-dependent modulation of those biases. Together, our simulations and results provide strong evidence that a formalized, dynamic framework directly linked to observable behavior can facilitate the theoretical integration of linguistic and non-linguistic spatial systems.

Processing structured data is a continuing challenge for connectionist models that aim at becoming a plausible explanation of human cognition. The recently proposed linear Recursive Auto-Associative Memory (RAAM) model was shown to have a much higher encoding capacity and not to be subject to overtraining compared to classical RAAM. We assess the effect of terminal encoding on the performance of linear RAAM in case of encoding trees of ternary semantic propositions and we show that the highest representation capacity is achieved with (sparse) binary WordNet-based codes, compared to (symbolic) neutral and to (distributed) word co-occurrence based codes. Only with WordNet codes the model could generalize to processing structures that contain known words at new syntactic positions or contain novel words, as long as these shared semantic features with the words from the training set.

Self-control can be defined as choosing a large delayed reward over a small immediate reward, while precommitment is the making of a choice with the specific aim of denying oneself future choices [10]. The self can be viewed as a goal directed hierarchical system, where goals are internally specified according to value systems that are developed through experience [14]. Given a situation, the presence of more than one established value system can give rise to interpersonal conflicts. Such conflicts can refer to self-control problems where people might attempt to overcome them by applying precommitment [1, 11]. A computational model of interpersonal conflict is proposed where we implement two spiking neural networks as two players, learning simultaneously but independently, competing in the Iterated Prisoner's Dilemma (IPD) game. Learning links behaviour to the synaptic level by reinforcing stochastic synaptic transmission [15]. An interpretation of the IPD is that it demonstrates interpersonal conflict [5]. It is possible, Kavka suggests, that such inner conflicts are resolved as if they were a result of strategic interaction among rational subagents [5]. The structure of the sub-agents' value systems is investigated with respect to the cooperative outcome of the game, which corresponds to the self controlled behaviour. The results seem to suggest that the degree of cooperation in the IPD depends on the structure of the payoff matrix. The relationship between precommitment behaviour and the value systems is also investigated and compared to our previous work [3]. In fact, with a technique resembling the precommitment effect, whereby the payoffs for the dilemma cases in the IPD payoff matrix are differentially biased (increased or decreased) [3], cooperation seems to be enhanced as the differential bias is increased.

According to an influential theory of cognitive control, conflict between competing responses is monitored and used to exert control over information processing. In this paper, I shy away from debates in the literature and ask whether there are other uses for monitored conflict. I present simulation results showing that conflict can be used to (1) produce retrospective confidence judgements, (2) dynamically adjust the response threshold, and (3) modulate stimulus-response learning. Although predictions need to be tested, the general conclusion is that conflict can be involved in metacognitive control.

In this paper we argue that computational issues like complexity, memory requirements and training time impose strong constraints on learning in any goal-oriented system. Along these constraints we derive a particular architecture that learns representations for optimizing plans e.g., trajectory planning. To comply with biological constraints as well, the resulting encoding mechanism is translated into a connectionist network. We argue that the goal-oriented framework implies distinct representations of place and direction in the hippocampal formation responsible for spatial navigation in mammals.

Computational modellers are not always explicit about their motivations for constructing models, nor are they always explicit about the theoretical implications of their models once constructed. Perhaps in part due to this, models have been criticised as "black-box" exercises which can play little or no role in scientific explanation. This paper argues that models are useful, and that the motivations for constructing computational models can be made clear by considering the roles that tautologies can play in the development of explanatory theories. From this, additionally, I propose that although there are diverse benefits of model building, only one class of benefits — those which relate to explanation — can provide justification for the activity.

This paper presents a localist multimodal neural network that uses Hebbian learning to acquire one-word child language from child directed speech (CDS) comprising multi-word utterances and queries in addition to one-word utterances. The model implements cross-situational learning between linguistic words used in child directed speech, the accompanying perceptual entities, conceptual relations and inferred communicative intentions. In 90 cases out of 117, the network successfully generates one-word utterances that may be viewed as being semantically equivalent to the CDS input used to train the network. The model also successfully emulates the one-word speech of a child in 12 out of 28 cases, despite its localist nature, thereby suggesting that Hebbian learning, as used in most models of cognitive development, is capable of cross-situational learning, a key component of multimodal temporal cognitive acquisition tasks, of which child language acquisition is one.

We present a neural-symbolic learning model of sentence production which displays strong semantic systematicity and recursive productivity. Using this model, we provide evidence for the data-driven learnability of complex yes/no-questions.

Classic connectionist models of reading have traditionally focused on English, a language with a quasi-regular (deep) relationship between orthography and phonology, and very little work has been carried out on more transparent (shallow) orthographies. This paper introduces a parallel distributed processing (PDP) model of reading for Italian. The model is successful in simulating a variety of behavioral effects such as the neighborhood effect and the morphological effect in nonword reading, previously accounted for by dual route architectures, and provides clear evidence that different grain sizes in the orthography to phonology mapping can be discovered even in a model trained with almost perfectly shallow stimuli.

Taking seriously neurobiological and psychological evidence on the constructivist, experience dependent nature of brain development, we present a constructivist neural network model which builds its architecture in response to the task of learning German (past participle) inflection. Our model captures developmental profiles, as well as healthy and impaired adult performance, because two complementary processing pathways develop from the interaction of the constructivist learning mechanism and the distributional properties of the inflectional paradigm. Instead of a regular/irregular dichotomy it suggests an emergent dissociation between verbs that are easy or hard to learn, thus obviating the need for in-built assumptions such as verb type specific processing mechanisms or knowledge of grammatical class. We focus on the German participle in order to demonstrate that the performance of the model, though based on associative learning mechanisms, does not depend on the existence of a dominant 'default' class, as has been claimed by proponents of the dual-mechanism camp within the continuing past tense debate.

For the last twenty years, many researchers interested in language acquisition have quantified the receptive and productive vocabulary of infants using CDIs – checklists of words filled in by the caregiver. While it is generally accepted that the caregiver can reliably say whether the infant knows and/or produces a given word, we lack an estimate for words that are not listed on CDI. In this study, we provide a mathematical model providing a link between CDI reports and a more plausible estimate of vocabulary size. The model is constrained by statistical data collected from a population of infants and is validated on a longitudinal study comparing diary report with CDI measures.

Performing conceptual tasks that do not involve overt sensory and motor processes, can nonetheless implicate sensory and motor regions of the brain. In models of "embodied" cognition, the sensory and motor brain regions are seen as integral to the representation of the concept. Alternatively, "disembodied" theories of concept representation assume that these activations are peripheral and epiphenomenal to the representation itself. We review three sources of data for embodied cognition – the activation of sensory and motor regions for conceptual tasks, the effect on conceptual task performance when motor areas are otherwise engaged, and behavioral influences on reading in patients with impaired sensory and motor areas. We show that such data is consistent with a connectionist model of embodied cognition, and discuss the sources of data that can distinguish between embodied and disembodied accounts.

Marchman's [1] framework for simulating sensitive periods in development was extended to investigate whether competition is a mechanism that might contribute to reductions in functional plasticity with age. Under this view, the ability to learn new behaviors is reduced when old established behaviors are unwilling to give up shared representational resources. The simulations supported this hypothesis, but indicated that a range of factors modulated competition effects: the similarity between old learning and new learning, the level of representational resources, the prevailing plasticity conditions within the system, the timing of introduction of new learning, and the complexity of the problem domain.

This paper explores the idea that auto-teaching neural networks with evolved self-supervision signals can lead to improved performance in dynamic environments where there is insufficient training data available within an individual's lifetime. Results are presented from a series of artificial life experiments which investigate whether, when, and how this approach can lead to performance enhancements, in a simple problem domain that captures season dependent foraging behaviour.

V1 receptive fields show different sensitivities to different scales suggesting spatial frequency coding of visual information. The purpose of the present paper is to determine whether such a spectral decomposition at a perceptual level would be efficient for non-linear categorization purposes. Specifically, we compare the advantage of providing an artificial neural network with this biologically plausible visual information vs. direct pixel information in a natural scene classification task. In order to keep information qualitatively constant in the two conditions, original images were downsampled in a way that preserved the same amount of data in the spatial frequency domain. Results show that a standard back-propagation neural network aimed at classifying visual images yields better performances when provided with data from Gabor receptive fields (simulating V1 biological neurons) than with direct pixel coding data. This result suggests that artificial and possibly biological neural systems would be better off using reduced Gabor filtered information in the spatial frequency (or spectral) domain rather than direct spatial information in the visual classification of natural scene images.

Increasingly cognitive scientists view prediction as central to cognition, from philosophical and psychological theories such as the enactive approach and sensorimotor perception [1-3], to computational neuroscience and modelling e.g. [4-7]. Indeed, as our experiments show, prediction can significantly increase the recognition of task relevant input features in models of cortical micro-columns (Echo State Networks) [8-11]. Here, feedback into the cortical micro-column is simply the output activity of single layer perceptrons trained to identify input features from the activity of the cortical micro-column. The surprising result is that while this feedback or prediction enhances recognition of features relevant to making that prediction, it consistently reduces performance at recognising non-prediction-task-relevant-features and hence provides a model of sustained inattentional blindness. This is confirmed in our computational model of a sustained inattentional blindness task demonstrating the role of feedback in successfully tracking the attended object, and how this can result in blindness to the presence of an unexpected object. The model therefore suggests that tuning input filters with predictions of what the sensory system should expect, may be the cause of inattentional blindness. While Simons and Chabris [12] note that "we perceive and remember only those objects and details that receive focused attention." p. 1059, our model suggests that the relevant information simply was not there to be attended.

The complex neural circuits found in fMRI studies of human attention were decomposed using a model of spiking neurons. The model for visual search over time and space (sSoTS) incorporates different synaptic components (NMDA, AMPA, GABA) and a frequency adaptation mechanism based on I_AHP current. This frequency adaptation current can act as a mechanism that suppresses the previously attended items. It has been shown [1] that when the passive process (frequency adaptation) is coupled with a process of active inhibition, new items can be successfully prioritized over time periods matching those found in psychological studies. In this study we use the model to decompose the neural regions mediating the processes of active attentional guidance, and the inhibition of distractors, in search. Activity related to excitatory guidance and inhibitory suppression was extracted from the model and related to different brain regions by using the synaptic activation from sSoTS's maps as regressors for brain activity derived from standard imaging analysis techniques (FSL). The results show that sSoTS pulls-apart discrete brain areas mediating excitatory attentional guidance and active distractor inhibition.

The following sections are included:

• AUTHOR INDEX