Neocognitron: a self organizing neural network model for a mechanism of pattern recognition unaffected by shift in position

Application of neural networks to the modelling of some constitutive laws

This study investigates the modelling of constitutive laws of materials by neural networks. Material behaviour is no longer represented mathematically but is described by neuronal modelling. The main aim is to build a neural network directly from experimental results (the learning phase). We give several examples of constitutive laws (Hooke, Sargin, etc.) using a backpropagation algorithm. Then we show that abilities of adjustment, memorisation and anticipation of neural networks permit us to develop a method of classification of constitutive laws.

Image-based object recognition in man, monkey and machine

Theories of visual object recognition must solve the problem of recognizing 3D objects given that perceivers only receive 2D patterns of light on their retinae. Recent findings from human psychophysics, neurophysiology and machine vision provide converging evidence for 'image-based' models in which objects are represented as collections of viewpoint-specific local features. This approach is contrasted with 'structural-description' models in which objects are represented as configurations of 3D volumes or parts. We then review recent behavioral results that address the biological plausibility of both approaches, a well as some of their computational advantages and limitations. We conclude that, although the image-based approach holds great promise, it has potential pitfalls that may be best overcome by including structural information. Thus, the most viable model of object recognition may be one that incorporates the most appealing aspects of both image-based and structural description theories.

Evidence accumulation in cell populations responsive to faces: an account of generalisation of recognition without mental transformations

In this paper we analyse the time course of neuronal activity in temporal cortex to the sight of the head and body. Previous studies have already demonstrated the impact of view, orientation and part occlusion on individual cells. We consider the cells as a population providing evidence in the form of neuronal activity for perceptual decisions related to recognition. The time course on neural responses to stimuli provides an explanation of the variation in speed of recognition across different viewing circumstances that is seen in behavioural experiments. A simple unifying explanation of the behavioural effects is that the speed of recognition of an object depends on the rate of accumulation of activity from neurones selective for the object, evoked by a particular viewing circumstance. This in turn depends on the extent that the object has been seen previously under the particular circumstance. For any familiar object, more cells will be tuned to the configuration of the object's features present in the view or views most frequently experienced. Therefore, activity amongst the population of cells selective for the object's appearance will accumulate more slowly when the object is seen in an unusual view, orientation or size. This accounts for the increased time to recognise rotated views without the need to postulate 'mental rotation' or 'transformations' of novel views to align with neural representations of familiar views.

Neural processing in the subsecond time range in the temporal cortex

The hypothesis that cortical processing of the millisecond time range is performed by latency competition between the first spikes produced by neuronal populations is analyzed. First, theorems that describe how the mechanism of latency competition works in a model cortex are presented. The model is a sequence of cortical areas, each of which is an array of neuronal populations that laterally inhibit each other. Model neurons are integrate-and-fire neurons. Second, the model is applied to the ventral pathway of the temporal lobe, and neuronal activity of the superior temporal sulcus of the monkey is reproduced with the model pathway. It consists of seven areas: V1, V2/V3, V4, PIT, CIT, AIT, and STPa. Neural activity predicted with the model is compared with empirical data. There are four main results: (1) Neural responses of the area STPa of the model showed the same fast discrimination between stimuli that the corresponding responses of the monkey did: both were significant within 5 ms of the response onset. (2) The hypothesis requires that the response latency of cortical neurons should be shorter for stronger responses. This requirement was verified by both the model simulation and the empirical data. (3) The model reproduced fast discrimination even when spontaneous random firing of 9 Hz was introduced to all the cells. This suggests that the latency competition performed by neuronal populations is robust. (4) After the first few competitions, the mechanism of latency competition always detected the strongest of input activations with different latencies.

Limited translation invariance of human visual pattern recognition

Visual object recognition is considered to be largely translation invariant. An earlier study (Foster & Kahn, 1985), however, has indicated that recognition of complex novel stimuli is partially specific to location in the visual field: It is significantly easier to determine the identity of two briefly displayed random patterns if both stimuli are presented at the same, rather than at different, locations. In a series of same/different discrimination tasks, we characterize the processes underlying this "displacement effect": Horizontal and vertical translations are equally effective in reducing performance. Making the task more difficult by increasing pattern similarity leads to even higher positional specificity. The displacement effect disappears after rotation or contrast reversal of the patterns, indicating that positional specificity depends on relatively low levels of processing. Control experiments rule out explanations that are independent of visual pattern memory, such as spatial attention, eye movements, or retinal afterimages. Positional specificity of recognition is found only for same trials. Our results demonstrate that position invariance, a widely acknowledged property of the human visual system, is limited to specific experimental conditions. Normalization models involving mental shifts of an early visual representation or of a window of attention cannot easily account for these findings.

Consciousness in Neural Networks?

A combined neurophysiological and computational approach is reviewed that leads to a proposal for how neural networks in the temporal cortical visual areas of primates could function to produce invariant object representation and identification. A similar approach is then reviewed which leads to a theory of how the hippocampus could rapidly store memories, especially episodic memories including spatial context, and how later recall of the information to the neocortex could occur. Third, it is argued that the visual and memory mechanisms described could operate without consciousness, and that a different type of processing is related to consciousness. It is suggested that the type of processing related to consciousness involves higher-order thoughts ("thoughts about thoughts"), and evolved to allow plans, formulated in a language, with many steps, to be corrected. It is suggested that it would feel like something to be a system that can think linguistically (using syntax) about its own thoughts, and that the subjective or phenomenal aspects of consciousness arise in this way. It is further suggested that "raw sensory feels" arise in evolution because once some types of processing feel like something by virtue of a system capable of higher-order thoughts, it is then parsimonious to postulate that sensory and related processing, which has to be taken into account in that processing system, should feel like something. It is suggested that it is this type of processing, which must be implemented in neural networks, which is related to consciousness.

SCAN: A Scalable Model of Attentional Selection

This paper describes the SCAN (Signal Channelling Attentional Network) model, a scalable neural network model for attentional scanning. The building block of SCAN is a gating lattice, a sparsely-connected neural network defined as a special case of the Ising lattice from statistical mechanics. The process of spatial selection through covert attention is interpreted as a biological solution to the problem of translation-invariant pattern processing. In SCAN, a sequence of pattern translations combines active selection with translation-invariant processing. Selected patterns are channelled through a gating network, formed by a hierarchical fractal structure of gating lattices, and mapped onto an output window. We show how the incorporation of an expectation-generating classifier network (e.g. Carpenter and Grossberg's ART network) into SCAN allows attentional selection to be driven by expectation. Simulation studies show the SCAN model to be capable of attending and identifying object patterns that are part of a realistically sized natural image. Copyright 1997 Elsevier Science Ltd.

Invariant visual responses from attentional gain fields

Inferotemporal (IT) neurons exhibit a substantial degree of invariance with respect to translation of images used as visual stimuli. Through theoretical and computer-modeling methods, we show how translation-invariant receptive fields, like those of IT neurons, can be generated from the responses of V4 neurons if the effects of attention are taken into account. The model incorporates a recently reported form of attention-dependent gain modulation in V4 and produces IT receptive fields that shift so they are centered at the point where attention is directed. Receptive fields of variable, attention-controlled spatial scale are obtained when the mechanism is extended to scale-dependent attentional gain fields. The results indicate that gain modulation may play analogous roles in the dorsal and ventral visual pathways, generating transformations from retinal coordinates to body- and object-centered systems, respectively.

Invariant face and object recognition in the visual system

Neurophysiological evidence is described, showing that some neurons in the macaque temporal cortical visual areas have responses that are invariant with respect to the position, size and view of faces and objects, and that these neurons show rapid processing and rapid learning. A theory is then described of how such invariant representations may be produced in a hierarchically organized set of visual cortical areas with convergent connectivity. The theory proposes that neurons in these visual areas use a modified Hebb synaptic modification rule with a short-term memory trace to capture whatever can be captured at each stage that is invariant about objects as the object changes in retinal position, size, rotation and view. Simulations are then described which explore the operation of the architecture. The simulations show that such a processing system can build invariant representations of objects.

Noise injection into inputs in sparsely connected Hopfield and winner-take-all neural networks

In this paper, we show that noise injection into inputs in unsupervised learning neural networks does not improve their performance as it does in supervised learning neural networks. Specifically, we show that training noise degrades the classification ability of a sparsely connected version of the Hopfield neural network, whereas the performance of a sparsely connected winner-take-all neural network does not depend on the injected training noise.

HAVNET: A New Neural Network Architecture for Pattern Recognition

A new artificial neural network architecture, specifically designed for two-dimensional binary pattern recognition, is introduced. The network employs a unique similarity metric, based on the Hausdorff distance, to determine the degree of match between an input pattern and a learned representation. Use of this metric in the network leads to behaviour that is more consistent with human performance than that generated by similarity metrics currently in use in other artificial neural networks. A detailed description of the architecture, the learning equations, and the recall equations for the network are presented. An extension of the network is also described in which each class of learned objects is represented by multiple two-dimensional aspects. This extension greatly increases the utility of the network for tasks like character recognition and three-dimensional vision. The network is employed on an example pattern recognition task to demonstrate its application, with very good results. Copyright 1996 Elsevier Science Ltd.

Using Spatio-temporal Correlations to Learn Invariant Object Recognition

A competitive network is described which learns to classify objects on the basis of temporal as well as spatial correlations. This is achieved by using a Hebb-like learning rule which is dependent upon prior as well as current neural activity. The rule is shown to be capable of outperforming a supervised rule on the cross validation test of an invariant character recognition task, given a relatively small training set. It is also shown to outperform the supervised version of Fukushima's Neocognitron ([Fukushima, 1980]), on a larger training set. Copyright 1996 Elsevier Science Ltd.

The Role of Neuronal Synchronization in Response Selection: A Biologically Plausible Theory of Structured Representations in the Visual Cortex

Recent experimental results in the visual cortex of cats and monkeys have suggested an important role for synchronization of neuronal activity on a millisecond time scale. Synchronization has been found to occur selectively between neuronal responses to related image components. This suggests that not only the firing rates of neurons but also the relative timing of their action potentials is used as a coding dimension. Thus, a powerful relational code would be available, in addition to the rate code, for the representation of perceptual objects. This could alleviate difficulties in the simultaneous representation of multiple objects. In this article we present a set of theoretical arguments and predictions concerning the mechanisms that could group neurons responding to related image components into coherently active aggregates. Synchrony is likely to be mediated by synchronizing connections; we introduce the concept of an interaction skeleton to refer to the subset of synchronizing connections that are rendered effective by a particular stimulus configuration. If the image is segmented into objects, these objects can typically be segmented further into their constituent parts. The synchronization behavior of neurons that represent the various image components may accurately reflect this hierarchical clustering. We propose that the range of synchronizing interactions is a dynamic parameter of the cortical network, so that the grain of the resultant grouping process may be adapted to the actual behavioral requirements.

It can be argued that different aspects of purposeful behavior rely on separable processes by which sensory input is transformed into adjustments of motor activity. Indeed, neurophysiological evidence has suggested separate processing streams originating in the primary visual cortex for object identification and sensorimotor coordination. However, such a separation calls for a mechanism that avoids interference effects in the presence of multiple objects, or when multiple motor programs are simultaneously prepared. In this article we suggest that synchronization between responses of neurons in both the visual cortex and in areas that are involved in response selection and execution might allow for a selective routing of sensory information to the appropriate motor program.

Parallel optical-feature extraction by use of rotationally multiplexed holograms

We propose a novel rotational (peristrophic) multiplexing method of hologram recording for parallel opticalfeature extraction and report basic experimental results. The features to be extracted are line-segment orientations separated by 30 degrees . The extracted features can be used for f lexible pattern recognition.

A multiscale dynamic routing circuit for forming size- and position-invariant object representations

We describe a neural model for forming size- and position-invariant representations of visual objects. The model is based on a previously proposed dynamic routing circuit that remaps selected portions of an input array into an object-centered reference frame. Here, we show how a multiscale representation may be incorporated at the input stage of the model, and we describe the control architecture and dynamics for a hierarchical, multistage routing circuit. Specific neurobiological substrates and mechanisms for the model are proposed, and a number of testable predictions are described.

Modeling of component failure in neural networks for robustness evaluation: an application to object extraction

The robustness of neural network (NN) based information processing systems with respect to component failure (damaging of nodes/links) is studied. The damaging/component failure process has been modeled as a Poisson process. To choose the instants or moments of damaging, statistical sampling technique is used. The nodes/links to be damaged are determined randomly. As an illustration, the model is implemented and tested on different object extraction algorithms employing Hopfield's associative memory model, Gibbs random fields, and a self-organizing multilayer neural network. The performance of these algorithms is evaluated in terms of percentage of pixels correctly classified under different noisy environments and different degrees and sequences of damaging. The deterioration in the output is seen to be very small even when a large number of nodes/links are damaged.

Modeling perceptual learning with multiple interacting elements: a neural network model describing early visual perceptual learning

We introduce a neural network model of an early visual cortical area, in order to understand better results of psychophysical experiments concerning perceptual learning during odd element (pop-out) detection tasks (Ahissar and Hochstein, 1993, 1994a). The model describes a network, composed of orientation selective units, arranged in a hypercolumn structure, with receptive field properties modeled from real monkey neurons. Odd element detection is a final pattern of activity with one (or a few) salient units active. The learning algorithm used was the Associative reward-penalty (Ar-p) algorithm of reinforcement learning (Barto and Anandan, 1985), following physiological data indicating the role of supervision in cortical plasticity. Simulations show that network performance improves dramatically as the weights of inter-unit connections reach a balance between lateral iso-orientation inhibition, and facilitation from neighboring neurons with different preferred orientations. The network is able to learn even from chance performance, and in the presence of a large amount of noise in the response function. As additional tests of the model, we conducted experiments with human subjects in order to examine learning strategy and test model predictions.

ID number recognition of X-ray films by a neural network

Results of ID number recognition of X-ray films using the neuro-computing technique implemented on a conventional engineering workstation are reported. The neural network was trained to identify the ten digits of arabic numerals and two roman letters printed on X-ray films. We used 22 sheets of films for the training set and 23 sheets for the test set. Each sheet of film contains 25 characters. A 99.5% recognition rate on the character base was achieved.