Connectionist Models of Cognition and Perception

The following sections are included:

• PREFACE

• CONTENTS

This paper describes how Cell Assemblies (CAs) can be used to model low level psychological phenomena. There is a brief description of CAs and our CANT simulator to ground the discussion. CAs are used to categorise and disambiguate inputs. We describe an simulation that shows CAs can be used to describe pattern disambiguation. It proposes simulations with CAs that can lead to a neurological explanation for timing of categorisation and disambiguation. CAs can also be used to explain priming, and simulations to explore priming at a neural level are proposed.

In the brain, knowledge must be represented in separate maps, such as location, identity, rank order, and procedures. Solutions to the binding problem based on 'conjunctive nodes' and specific temporal codes are reviewed in the light of mechanisms of self-organization. A conceptual network will be discussed in which memory traces are carried by cell-assemblies with a critical threshold: when sufficient neurons are externally activated the overall activation level in the cell-assembly rises to its maximum by itself. Propagation of oscillatory excitation loops can take place above the critical threshold - the corresponding memory traces are then said to be in short-term memory - or below the critical threshold - the involved memory traces are then in a state of priming.

The bindings between the various maps at the neural level are assumed to exist in the form of temporary resonances between excitation patterns in cell-assemblies or in a spatial map. Two patterns are bound because they are active simultaneously and because they are activated from a common subnetwork that represents the momentary context in which the network is functioning.

Different forms of binding are discussed, illustrating the binding of location and identity and that of rank order and identity. A computer simulation and an experiment in relation with the former is described.

Few people would disagree that the human brain is modular, but there is less agreement on the reasons why it has evolved to be like that. Recently I re-examined the Rueckl, Cave & Kosslyn study⁹ which demonstrated the advantages of having a modular architecture in neural network models of a simplified version of the "what" and "where" vision tasks. Explicit evolutionary simulations confirmed that the advantage can cause modularity to evolve, but also demonstrated that simply changing the learning cost function produced a system that learnt even better than before, and in which modularity did not evolve. In this paper I attempt to find a more robust characterisation of the evolution of modularity in terms of gated sub-networks (i.e. mixtures of expert networks). Once again, a careful analysis of a systematic series of explicit evolutionary simulations indicates that drawing reliable conclusions in this area is not as straightforward as it might at first appear.

This work moves from the general hypothesis that action influences knowledge formation, and that the way we organise our knowledge reflects action patterns [7]. The traditional assumption in the categorisation literature is that categories are organised on the basis of perceptual similarity among their members. But much evidence shows that, when we need to perform an action, we can group objects which are perceptually dissimilar. Many studies have shown that we are able to flexibly organise and create new categories of objects on the basis of more or less contingent goals [2,3].

We present some simulations in which neural networks are trained using a genetic algorithm to move a 2-segment arm and press one of two buttons in response to each of 4 stimuli. The neural networks are required to group the stimuli, by pressing the same button, in 2 categories which, depending on the particular task (which is encoded in a set of additional input units), may be formed by perceptually very similar, moderately similar, or different objects.

We find that task information overrides perceptual information, that is, the internal representations of neural networks tend to reflect the current task and not the perceptual similarity between the objects. However, neural networks tend to form action-based categories more easily (e.g. in fewer generations) when perception and action are congruent (perceptually similar objects must be responded to by pressing the same button) than when they are not congruent (perceptually similar objects must be responded to by pressing different buttons). We also find that at hidden layers nearer the sensory input, where task information still has not arrived, internal representations continue to reflect perceptual information.

Young infants exhibit intriguing asymmetries in the exclusivity of categories formed on the basis of visually presented stimuli. For instance, infants who have previously seen a series of cats show a surge of interest when looking at dogs, this being interpreted as dogs being perceived as novel. On the other hand, infants previously exposed to dogs do not exhibit such an increased interest for cats. Recently, researchers have used simple autoencoders to account for these effects. Their hypothesis was that the asymmetry effect is caused by the smaller variances of cats' features and an inclusion of the values of the cats' features in the range of dogs' values. They predicted, and obtained, a reversal of asymmetry by reversing dog-cat variances, thereby inversing the inclusion relationship (i.e. dogs are now included in the category of cats). This reversal reinforces their hypothesis. We will examine the explanatory power of this model by investigating in greater detail the ways by which autoencoders exhibit such an asymmetry effect. We analyze the predictions made by a linear Principal Components Analysis. We examine the autoencoder's hidden-unit activation levels and, finally, we emphasize various factors that affect generalization capacities and may play key roles in the observed asymmetry effect.

The human ability to process images and understand what they mean in order to solve a problem holds an important clue to how the human thought process works. This suggests a representation which orientates itself on the perception of the world. But, is such a representation also suitable for modeling problem solving, a domain which is traditionally reserved for the symbolic approach? An important research field in artificial intelligence is the usage of heuristics to speed up the search. It is not easy to define heuristic functions, as there is no rule which says how to do this. Another approach to the view of heuristic functions is shown in this work: heuristic functions which result from the manner of description of knowledge, namely by perception orientated representation. Examples of such representation are two dimensional binary sketches, or numbers represented by bars at certain positions which can overlap. Experiments with a neural production system with this kind of representation are performed in the 8-puzzle and the block world domain.

Animate goal-directedness is characterised by plastic and persistent action where paths are continually shifted so that they project to the goal. The internal organisation that goal direction induces underpins the cognitive perception-action loop of intention.

Such organisation also provides a basis for more powerful artificial neural networks. Artificial neural search mainly employs forms of gradient descent, a mechanism based on inanimate dynamics. Consequently, re-initialisation and multiple runs have to be used because the likelihood of success on any single run is limited due to local minima, shallow gradients and misleading gradients pointing away from the goal.

Plastic and persistent search based on animate dynamics may be more successful, particularly with sequential problems. In these cases, continuous progress through a sequence of goals is required first-time without re-initialisation thus rendering gradient descent solution ineffective. Plastic persistence on the other hand is ideally suited to succeeding for these problems since its redirection ability means that no re-initialisation is required.

The paper describes a project for investigating the plastic and persistent path types found in human smooth goal-directed behaviour through measurement with a view to incorporating these into artificial heuristic search using neural networks. The design of future experiments is described for enabling this measurement to take place.

Habituation paradigms have become popular tools to investigate early cognitive abilities in infants. However, the level of interpretation that should be used in explaining habituation phenomena has been the object of much debate in recent years. Several neural network models of infant habituation phenomena have been proposed, and some of these are explicitly aimed at this issue. We argue that none of the previous models offer either a necessary or sufficient account of infant habituation. We look at the constraints from the behavioural and neural sciences that bound interpretations of habituation and familiarisation tasks. The main features identified at the neural level are selective inhibition in the hippocampal system, and a dynamic interaction between these subcortical processes and cortical regions. At the behavioural level, proper models should capture the temporal unfolding of responses, their exponential decrease, a shift from familiarity preference to novelty preference, the habituation of novelty preference, and the ability to discriminate between habituated stimuli. A neural network model built around these constraints is presented.

Habituation is the fundamental property that the efficiency of neural cells diminishes over repeated stimulation. A model for how habituation influences episodic tests is suggested. During encoding or presentation of stimuli, habituation causes changes in activity, which influences the degree of synaptic plasticity. Fast habituation occurs during the short time intervals of presentation of single items and slow habituation during long time intervals over the presentation of a list if items. Item habitation causes more efficient encoding at the beginning of the presentation of an item, whereas list habituation improves encoding at the beginning of the lists as seen in the primacy effect. The rate of habitation increases with the frequency of items. The habituation model makes specific predictions of how performance depends on study time, serial position, and frequency of items. The model is supported by a set of experimental data.

We present an activation-based computational model of immediate item memory, which is proposed to underlie activation based processes within the pre-frontal cortex (mediating primary memory) that trigger episodic learning processes in the medial temporal cortex (secondary memory). We show that the model is able to capture a range of basic phenomena such as Brown-Peterson forgetting functions and serial position functions in cued and free-recall. The model makes unique predictions for presentation rate and list length effects, which were tested (and supported) in subsequent experiments.

A distributed connectionist model of spiking neurons (INFERNET) is used to simulate various aspects of Short Term Memory. In INFERNET, short term memory is the transient activation of long term memory elements. This single store model has a human-like performance in short term memory span tasks, but also displays serial position effects, similarity effects, and double dissociation between short and long term memory which are considered as the main psychological arguments in favor of the multiple-store model.

Variations in face shapes were analysed using principal and independent components analysis. After removing components that code mainly for head movement, face images were produced that depicted extremes at each end of the range of seven components. These faces were rated on a number of characteristics, such as masculinity and intelligence. Several of the components showed significant differences on the ratings but there was no clear evidence for an advantage for either method of analysis.

Specific cells in the visual cortex serve to detect vertices, such as crossing lines and edges, or to construct illusory contours in the case of occlusion, as occurs with the Kanisza triangle. Computational models of these cells have already been published. Yet other cells serve to detect periodic gratings (textures) or isolated bars. Computational models of the latter cells have also been published, but they lack a precise localisation and the bar cell model is not really a bar detector because it detects anything that is not a periodic grating. We present improved models for grating and bar cells. These models employ a common frontend that consists of retinal ON and OFF channels (isotropic DOG filters) in combination with shunting networks for a contrast normalisation, followed by anisotropic filtering (Gabor filters as a model for simple cells) and an edge sharpening. The resulting ON and OFF responses are then used with different neural grouping operators (dendritic combination fields) to detect either periodic gratings or individual bars. The models are very selective in terms of grating frequency, bar width and orientation, and result in a very precise boundary localisation. A complete cell assembly covers all frequencies, widths and orientations at each retinotopic position. Such an assembly, together with other functional units, can provide a computational framework for explaining at least low-level cognitive effects.

Recent work suggests that visual information can be encoded by the activities of small proportions of many available units [1]. This can be seen as energetically efficient [2] as few outputs are strongly active at any time. An image coding model is presented to examine the effect of an efficient use of synapses rather than outputs. The model has two main parameters, the input:output ratio r, and a total weight energy budget k. It was found that for a given number of output units there is a weight budget that leads to optimal efficiency, balancing network performance and energy expenditure. More outputs can achieve higher efficiency, but this requires a higher budget. The resulting filters have center surround organization with strong quantitative similarities to retinal ganglion cell receptive fields. This work contributes to a common energy efficiency interpretation of sensory and cortical level processing in the visual system.

Cognition of abstract visual shape properties and spatial relations go beyond object recognition. Ullman proposed visual routines¹⁶ as the mechanism for flexible creation of different abstract visual features, each routine being a sequence composed of more basic visual elemental operations. We re-examine visual routines, inspired by recent neurophysiological advances (Roelfsema et al., 2000) and computational models of the primate visual system and propose possible underlying mechanisms for implementing elemental operations upon which visual routines are based. Whereas previous approaches to visual routines emphasized boundary contour processing or filling-in of regions between contours, we focus on medial axes and image segmentation in addition to boundary contours for a more inclusive understanding of visual routines. We show, at least conceptually, that all visual routines can be interpreted in terms of image segments. This framework suits the extraction of various abstract shape properties necessary to produce visual routines which agree with human visual perception in tasks such as discrimination of spirals, texture segregation, size illusions and more.

What determines where we look? Eye position data from subjects viewing a range of images of natural scenes are used to investigate the positions that are fixated by the eyes over the course of several seconds of viewing. This behavioural data suggests that both exogenous and endogenous factors are involved in the targeting of saccades. Fixation positions are then modeled with respect to low-level image features by the construction of saliency maps for 4 different regularities, each at 3 spatial scales. We find that contrast and edge-content saliency maps, particularly at the fine spatial scales, are able to account for some of the behavioural data, but chromaticity and luminance do not appear to be involved in determining where we look. Over the course of viewing, the behavioural data suggests a change in the relative contributions of exogeny and endogeny in eye movement control.

Masked priming experiments have shown that perceptuo-motor linkages can be made below the threshold of conscious experience. Notable amongst these experiments is work by Eimer and Schlaghecken that has shown that negative compatibility effects can be obtained, whereby behavioural costs are incurred when prime and target are compatible. Negative compatibility is suggestive of inhibitory mechanisms; a theory that is supported by lateralized readiness recordings of motor cortex activation. This paper develops a neural network model of masked priming that yields such negative compatibility. The key mechanisms that facilitate this outcome are (lateral inhibition based) competition between response alternatives and (opponent processing based) threshold gated direct response suppression. The main result of our model is the generation of response readiness profiles commensurate with the lateralized readiness potentials recorded from humans.

In this paper we examine the core of a recently proposed model of the entorhinal-hippocampal loop (EHL). The core model is built on pure information theoretic principles. It accounts for the most characteristic features of the EHL including long-term memory (LTM) formation. Here we argue that the core model, which performs novelty detection, provides correct temporal ordering for learning. Surprisingly, as we examine the temporal characteristics of the model, the experienced dynamics can be interpreted as the perceptual priming phenomenon. Computational results support the hypothesis that there might be a strong correlation between perceptual priming and repetition suppression and this correlation is a direct consequence of the temporal ordering in forming the LTM. We argue also that a relatively simple and coherent explanation of priming positions our model as a possible prototype for neocortical layers.

This paper aims at identifying the regions of interest in natural scenes. These regions have been defined by a behavioural measure of eye movement and by a model of saliency map constructed in a biologically plausible manner.

The saliency map codes the local region of interest in terms of signal properties such as contrast, orientation, colour, curvature etc. In our approach, pictures are processed using a retinal model, simulating the parvocellular output of the retina. The result is then filtered by a bank of Gabor filters, in mutual interaction in order to lower noise, enhance contour, and sharpen filter selectivity.

Subjects' eye positions were recorded as they explored static black and white images in order to categorize these images. All fixations during one scene were averaged in order to make a density map coding the time spent for subjects on each pixel. Statistics were computed on the regions around the fixation point to evaluate an index of predictability of our saliency map. The saliency map and the density map select similar areas. Furthermore, statistics based on eye-selected regions show greater values than for randomly-selected ones.

The effect of social categorisation on optical illusions such as the Ebbinghaus illusion is an example of how social categorisation impinges on perceptual treatment of stimuli. The widespread explanation for this illusion relies on size contrast theory. Nevertheless, Stapel & Koomen's studies [28] have shown that increasing conceptual or social category similarity between central and surrounding elements enhances the illusion, suggesting that the perceptual processes involved occur at a later stage of processing. This effect diverges from a prediction based on Self Categorisation Theory, according to which increasing social category similarity decreases the illusion, since differences between elements belonging to the same category are minimised. We conceptually replicated Stapel & Koomen's study. Our results show an inverse pattern to theirs: while obtaining an illusion in all conditions, the illusion increased when category similarity decreased. Within a SCT perspective, we propose an explanation for the difference between these diverging results based on the social nature of the stimuli.

We present a neural network model that integrates between motor commands and their associated perceived sounds in speech production. The model is based on simple, neurobiologically plausible mechanisms. In modelling the development of vowels in an infant's babbling phase, perceptual and action prototypes develop concurrently. When the model is exposed to "external" vowels, the perceptual prototypes develop to coincide with the external sounds, leading to language-specific categorical perception. The model develops motor mirror neurons that are activated when their associated sounds are perceived, and it learns to imitate vowels based on its own production of these vowels. We suggest an extension of the model to account for visual mirror neurons.

This paper concerns the perception of syllable structure and the way this process allows us to form structural representations of syllables in memory. It is argued that bottom-up processes are required for the formation of phonological representations of nonwords in short-term memory, and of words in early childhood, where top-down knowledge concerning syllable structure cannot be assumed. A new model is presented using signal processing techniques to show that representations of syllable structure can be developed from acoustic information, without top-down input. The model is motivated by psycholinguistic data and theory, expressed at the level of a putative neural mechanism, and supported by recent neurophysiological and brain imaging data. It describes syllable structure not in terms of events (e.g., syllable boundaries) or phonological frames with discrete slots, but in terms of a continuously varying measure of syllable position: syllabic phase. The validity of the model is demonstrated using a corpus of spoken sentences from the TIMIT database. Implications of the model for theories of phonological memory and developmental dyslexia are discussed. On the basis of recent functional imaging and neurophysiological studies, the neural basis for the proposed mechanism is hypothesized to be in the posterior Superior Temporal Sulcus.

The present chapter reports an experimental study of the semantic basis of lexical selection in speech. In a picture-word interference study, subjects named pictured objects while ignoring semantically related or unrelated distracters. It was found that the related distracters inhibited picture naming in comparison to the unrelated distracters. Indeed, the extent to which a related distracter elicited inhibition was predicted by the semantic similarity of distracter and target picture lexical concepts. Semantic inhibition is thought to show the effect of distracter presentation on the competition for lexical selection between candidate words in conceptually-driven speech. The results of the experiment suggest that such competition is more intense for more similar words. Measures of semantic similarity derived from lexical co-occurrence did not usefully predict observed semantic inhibition. The rated similarity of stimulus pairs was, however found to be useful in predicting inhibition. The results are discussed in relation to current models of lexical selection and of semantic space.

The following sections are included:

• AUTHOR INDEX