Skip to main content

Über dieses Buch

This is an exciting time. The study of neural networks is enjoying a great renaissance, both in computational neuroscience - the development of information processing models of living brains - and in neural computing - the use of neurally inspired concepts in the construction of "intelligent" machines. Thus the title of this volume, Dynamic Interactions in Neural Networks: Models and Data can be given two interpretations. We present models and data on the dynamic interactions occurring in the brain, and we also exhibit the dynamic interactions between research in computational neuroscience and in neural computing, as scientists seek to find common principles that may guide us in the understanding of our own brains and in the design of artificial neural networks. In fact, the book title has yet a third interpretation. It is based on the U. S. -Japan Seminar on "Competition and Cooperation in Neural Nets" which we organized at the University of Southern California, Los Angeles, May 18-22, 1987, and is thus the record of interaction of scientists on both sides of the Pacific in advancing the frontiers of this dynamic, re-born field. The book focuses on three major aspects of neural network function: learning, perception, and action. More specifically, the chapters are grouped under three headings: "Development and Learning in Adaptive Networks," "Visual Function", and "Motor Control and the Cerebellum.



Dynamic Interactions in Neural Networks: An Introductory Perspective

Dynamic Interaction in Neural Networks: An Introductory Perspective

It is the purpose of this introduction to briefly review the papers in each of the three parts of the volume, and then conclude with a brief unifying perspective.
Michael A. Arbib

Development and Learning in Adaptive Networks


Dynamical Stability of Formation of Cortical Maps

A cortical map is a localized neural representation of the signals in the outer world. A rough map is formed under the guidance of genetic information at the initial stage of development, but it is modified and refined further by self-organization. The persent paper gives a mathematical theory of formation of a cortical map by self-organization. The theory treats both dynamic of excitation patterns and dynamics of self-organization in a neural field. This not only explains the resolution and amplification properties of a cortical map, but elucidates the dynamical stability of such a map. This explains the emergence of a microcolumnar or mosaic structure in the cerebrum.
Shun-ichi Amari

Visual Plasticity in the Auditory Pathway: Visual Inputs Induced into Auditory Thalamus and Cortex Illustrate Principles of Adaptive Organization in Sensory Systems

We have induced, by appropriate surgery in newborn ferrets, retinal projections into the medial geniculate nucleus, the principal auditory thalamic nucleus. In operated animals studied as adults, retinal ganglion cells that give rise to the projection have small and medium sized somata and heterogeneous dendrite morphologies. Each retina projects to the auditory thalamus in patchy fashion. Various nuclei in auditory thalamus project normally to auditory cortex. Visual cells in auditory thalamus have circular receptive fields and receive input from slowly conducting afferents characteristic of retinal W cells. Many visual cells in primary auditory cortex have oriented receptive fields that resemble those of complex cells in striate cortex. Primary auditory cortex also contains a two dimensional visual field map. Our results carry several implications for sensory cortical function. A parsimonious explanation for the visual receptive field properties in auditory cortex is that sensory cortex carries out certain stereotypical transformations on input regardless of modality. The response features of visual cells and the two dimensional visual field map in primary auditory cortex appear to be products of adaptive organization arising from a highly divergent thalamocortical projection characteristic of the auditory system.
Mriganka Sur

The Hippocampus and the Control of Information Storage in the Brain

The present chapter assumes that the amount of information stored in the brain at a given moment is proportional to the mismatch between internally predicted events and the actual events ocurring in the external world. We have proposed ([29], [30]) that the hippocampus is involved in the computation of the “aggregate prediction” of ongoing events. This prediction is compared with information from the external world in order to determine the amount of information to be stored in the brain.
According to the “aggregate prediction” hypothesis (a) the effect of hippocampal lesions (HL) is an impairment in the integration of the aggregate prediction used to compute attentional variables (b) the effect of the induction of hippocampal long-term potentiation (LTP) is an increase in the value of the aggregate prediction by way of increasing the value of CS-CS associations and (c) that neural activity in hippocampus is proportional to the instantaneous value of the aggregate prediction. In addition, the present chapter introduces the hypothesis that medial septum activity is proportional to the sum of the values of different attentional variables.
The present chapter presents computer simulations for delay conditioning, conditioned inhibition, extinction, latent inhibition, and blocking for normal and HL cases. The “aggregate prediction” hypothesis proved capable of simulating most, but not all, experimental data regarding hippocampal manipulations in the rabbit nictitating membrane response preparation.
Nestor A. Schmajuk

A Memory with Cognitive Ability

The cognition is based on classification processes. In every stage of computation or information processing, a sub-unit in each site of the brain will also classify whether or not the input signal to that unit is identifiable in its part.
Shigeru Shinomoto

Feature Handling in Learning Algorithms

TLU (Threshold Logic Unit) representation and training provides a simplified formal model of neuron-like computation. Based on this abstract model, various formal properties and acceleration techniques are considered, the results of which appear relevant to observed biological phenomina.
S. E. Hampson, D. J. Volper

Self-Organizing Neural Network with the Mechanism of Feedback Information Processing

Several neural network models, in which feedback signal effectively acts on their functions, are proposed.
A multilayered network which has not only feedforward connections from the deepest-layer cells to front-layer cells, is proposed. The feedback connections, as well as the conventional feedforward connections, are self-organized. After completion of the self-organization, even though an imperfect or an ambiguous pattern is presented, the response of the network usually converges to that for one of the learning patterns. It may be seen that the network has characteristics quite similar to the associative recall in the human memory.
A rule for the modification of connections is proposed, suggested by the hypothesis that the growth of connections is controlled by feedback information from postsynaptic cells. Even if a new pattern resembling to one of the learning patterns with which the network has been organized, is presented the network is capable to be self-organized again, and a cell in the deepest layer comes to acquire a selective responsiveness to the new pattern. This model shows a characteristic closely resembling to that of human being, such as the ability to adapt flexibly to different new environments.
A model which has modifiable inhibitory feedback connections between the cells of adjoining layers, is proposed. If a feature-extracting cell is excited by a familiar pattern, the cell immediately feeds back inhibitory signals to its presynaptic cells. On the other hand, the feature-extracting cell does not respond to an unfamiliar pattern, and the responses from its its presynaptic cells are therefore not suppressed. In the network, connections from cells yielding a large sustained output are reinforced. Since familiar features do not elicit a sustained response from the cells of the network, only circuits detecting novel features develop. The network therefore quickly acquires favorable pattern-selectivity.
Sei Miyake, Kunihiko Fukushima

Visual Function


Interacting Subsystems for Depth Perception and Detour Behavior

Where many models of depth perception focus on the processing of disparity cues alone, we here present two models of depth perception, the Cue Interaction model and the Prey-Localization Model, which involve cooperative computation using both disparity and accomodation as sources of depth information. We then introduce models of detour behavior in which such depth schemas can function as subsystems.
Michael A. Arbib

Role of Basal Ganglia in Initiation of Voluntary Movements

A motor system called the basal ganglia facilitates movement initiation by removing its powerful inhibition on other motor areas. It may also facilitate activity in the cerebral cortex with disinhibition and ensure sequential processing of motor signals.
Okihide Hikosaka

Neural Mechanisms of Attention in Extrastriate Cortex of Monkeys

Neuronal recordings in extrastriate cortex of awake monkeys have shown that sensory processing is under the control of selective attention. Selective attention serves to remove irrelevant information from the receptive fields of extrastriate neurons and sharpen their selectivity for visual features. These effects of attention may explain both why we have little awareness of unattended stimuli, and why our resolution of spatial location and visual features is improved inside the focus of attention.
Robert Desimone, Jeffrey Moran, Hedva Spitzer

Neuronal Representation of Pictorial Working Memory in the Primate Temporal Cortex

It has been proposed that visual memory traces are located in the temporal lobes of the cerebral cortex, as electric stimulation of this area in humans results in recall of imagery1. Lesions in this area also affect recognition of an object after a delay in both humans2,3 and monkeys4–7, indicating a role in working memory of images8. Single-unit recordings from the temporal cortex have shown that some neurons continue to fire when one of two or four colors are to be remembered temporarily9. However, neuronal responses selective to specific complex objects10–18, including hands10,13 and faces13,16,17 cease soon after the offset of stimulus presentation10–18. These results left it open whether any of these neurons could serve memory of the object. We have recently found a group of shape-selective neurons in an anterior ventral part of the temporal cortex of monkeys that exhibited sustained activity during the delay period of a visual working memory task19,20. The activity was highly selective for the pictorial information to be memorized and was independent of the physical attributes such as size, orientation, color or position of the object. These observations indicate that the delay activity represents the working memory of categorized percept of a picture. This article discusses the implications of these findings in the cognitive neuroscience.
Yasushi Miyashita

Motor Control and the Cerebellum


Hierarchical Learning of Voluntary Movement by Cerebellum and Sensory Association Cortex

In earlier papers, we have proposed the feedback-error-learning of inverse dynamics model of the musculoskeletal system as heterosynaptic learning scheme in the cerebrocerebellum and the parvocellular part of the red nucleus system, and the iterative learning in the parietal association cortex. In this paper, we applied hierarchical arrangement of these two neural network models to learning trajectory control of an industrial robotic manipulator. We found that the hierarchical arrangement of the cerebellar and cerebral neural networks not only increased control stability but also dramatically improved accuracy of control and reduced required learning time.
Mitsuo Kawato, Michiaki Isobe, Ryoji Suzuki

A Model for Oblique Saccade Generation and Adaptation

In 1975, D.A. Robinson developed a model to illustrate the generation of saccadic eye movements. This model has been a potential working hypothesis for the study of saccades. It explains that a saccadic control system is essentially a position-servo system with a high-gain forward path. The characteristics include the following two points: 1) A neural integrator, which eliminates steady-state positional error in the servo system, is involved in the concept of a “final common path”. The path compensates for the dynamics of the oculomotor system. 2) Pause neurons suppress the instability of the high-gain servo system by placing a deadband element in a high-gain forward path. In spite of logical completeness, the desirable position signals or feedback positional error signals cannot be found in the actual brain stem circuitry, as was expected from the model.
Masahiko Fujita

Cerebellar Mechanisms in the Adaptation of Vestibuloocular Reflex

A lot of efforts have been devoted to identify neuronal circuitry responsible for adaptation of the vestibuloocular reflex (VOR) and to build a control system model for the whole VOR system. However, a controversy still exists in evaluating the crucial roles of the cerebellum in the VOR adaptation, and it seems urgent to solve the controversy on a sound experimental basis. One clue for solution may be obtained from examination of involvment of the inferior olive in the VOR adaptation, as the inferior olive has been assumed to play a key role in learning mechanisms of the cerebellar cortex.
Yasushi Miyashita, Koichi Mori

A Kalman Filter Theory of the Cerebellum

A variety of evidence suggests that the cerebellum is directly involved in certain sensory tasks. The specific hypothesis developed in this article is that the cerebellum is a neural analog of a Kalman-Bucy filter, whose function is to estimate state variables of the motor system and of external dynamical systems.
Michael Paulin

Conditioning and the Cerebellum

Mathematical models of learning that describe the effects of training with no regard for motor output are incomplete and consequently difficult to represent within natural systems. This chapter summarizes our efforts to reconcile a model that describes real-time topographical features of the classically conditioned nictitating membrane response (NMR) of the rabbit with knowledge about the cerebellum, the brain region thought to be most crucial for NMR conditioning. The NMR is a protective response resulting from retraction of the eyeball and the passive sweeping of the NM over the eye. The conditioned NMR is a graded, adaptive response.
John W. Moore, Diana E. J. Blazis


Weitere Informationen