Search

In a previous paper a method was given by which the efferent activity of an idealized neural net could be calculated from a given afferent pattern. Those results are extended in the present paper. Conditions are given under which nets may be considered equivalent. Rules are given for the reduction or extension of a net to an equivalent net. A procedure is given for constructing a net which has the property of converting each of a given set of afferent activity patterns into its corresponding prescribed efferent activity pattern.

This paper describes signal-processing techniques for the recognition of speech phonemes by machine. An attempt has been made to employ, wherever useful, basic processing functions of the human auditory system. These basic functions include neural interconnections and the mechanical transfer functions of the receptor organs. The neural interconnections bave been simulated by the use of neural logic. The purpose of this paper is to describe the logic networks that have been developed for the abstraction of speech features.

A feasible pavement crack detection system plays an important role in evaluating the road condition and providing the necessary road maintenance. In this paper, a back propagation neural network (BPNN) is used to recognize pavement cracks from images. To improve the recognition accuracy of the BPNN, a complete framework of image processing is proposed including image preprocessing and crack information extraction. In this framework, the redundant image information is reduced as much as possible and two sets of feature parameters are constructed to classify the crack images. Then a BPNN is adopted to distinguish pavement images between linear and alligator cracks to acquire high recognition accuracy. Besides, the linear cracks can be further classified into transversal and longitudinal cracks according to the direction angle. Finally, the proposed method is evaluated on the data of 400 pavement images obtained by the Automatic Road Analyzer (ARAN) in Northern China and the results show that the proposed method seems to be a powerful tool for pavement crack recognition. The rates of correct classification for alligator, transversal and longitudinal cracks are 97.5%, 100% and 88.0%, respectively. Compared to some previous studies, the method proposed in this paper is effective for all three kinds of cracks and the results are also acceptable for engineering application.

A hypothetical neural network is presented to account for errors made by food-deprived albino-rats in a bar-pressing situation (Carlton, 1964). The rats were required to press alternately on two bars in order to activate a device that releases milk to them. When-ever an animal pressed consecutively on the same bar, no milk was released and the animal was scored as having committed an error. In the first part of the paper, a time-independent neural network, part of which is equivalent to the psychophysical discrimination network of H. D. Landahl (1938), is used to interpret the effects on the animals’ performances of the drugs amphetamine, scopolamine and atropine. Suggestions for further experiments are made on the basis of the initial form to the model. In the second part of the paper, certain parameters of the initial form of the model are assumed to be time-dependent and a further generalization occurs through the introduction of the interresponse time distribution. It is shown that, under specified conditions, the second form of the model reduces to the first. The time-dependent form of the model allows certain features to be discussed that could not be discussed in the time-independent form [e.g. P. Dew’s (1961) notion of a possible mode of action for a variety of the behavioral effects of amphetamine]. Furthermore, experiments of an essentially different type from those discussed in the first part can be proposed to aid in the development of a theory for this kind of behavioral situation.

This paper constitutes Part II of a series of papers on adaptive pattern recognition and its applications. It pertains to optimal, unsupervised learning, adaptive pattern recognition of "lumped" gaussian signals in white gaussian noise. Specifically, both deterministic decision directed learning as well as random decision directed learning algorithms for continuous data are obtained. It is shown that the supervised learning results [1], in particular the partition theorem are applicable in the directed learning approach to the unsupervised case [2].

Anatomical and physiological evidence is cited for the existence in the CNS of more or less discrete populations of interconnected neurons. These are given the term netlets. A model based on these observations is presented, in which it is assumed that the netlets are the fundamental building blocks out of which nets of considerable complexity may be as embled. The connectivity within each netlet is assumed to be random. Neuronal macrostates are defined in which the fractions of neurons active in each netlet are the dynamical variables. Thus the temporal and spatial fine structure of neuronal activity are considered to be of secondary significance and are disregarded. These assumptions bring about an enormous reduction in complexity. Thus calculations and computer simulation studies become possible for systems hitherto inaccessible to quantitative description. It is hoped that the features retained in the model play a sufficiently significant role in the functioning of real neural nets to make these results meaningful. The mathematical formalism and detailed numerical results appear in another paper of this issue ( Anninos, 1970 ). Some of these results are anticipated in this paper and their implications for our model are discussed. The study proceeds from a treatment of isolated probabilistic netlets to the dynamics of interacting netlets. Of particular interest are the conditions under which a netlet will go into sustained activity and the often extremely delicate control exerted by afferent excitatory or inhibitory biases. Hysteresis effects are common and may represent a type of short-term memory. A variety of neural functions are listed to which some of these mechanisms may be applied. Among these are the modulating effects of the brain stem reticular formation on cortical and spinal neuron populations and the “energizing” of cortical centers by spontaneous activity in sensory systems. Finally the concepts of netlet interaction are applied in conjunction with the principle of synaptic facilitation to information processing in the cortex. Examples given are sensory-sensory cortical conditioning and the formation of the classical conditioned reflex.

Perceptrons are self-organizing or adaptive systems proposed by Frank Rosenblatt as greatly simplified models for biological brains. The main objective is to begin to explain how the brain performs its functions in terms of its structural components. Consequently the methodology consists largely of investigating the behaviour of neural networks which, except for oversimplification, are not unreasonable models of the brain structure, and searching for non-trivial psychological behaviour. This is in contrast with the customary engineering approach of first deciding what function is to be performed and then designing a system to perform the desired function.A perceptron is a network consisting of ideal neurons, similar to those of McCulloch and Pitts, connected together more or less at random, subject to certain organizational constraints and laws of growth. In this paper we give a brief introduction to the subject and describe some of the results of the mathematical analysis of several such systems; namely 'simple' perceptrons and 'four layer series-coupled' perceptrons.

Brindley (1967, 1969, 1972) has discussed nets of several types of formal neurons, many of whose functions are modifiable by their own input stimuli. Because Brindley's results are widely referred to, for example Marr (1970, 1971) and include some of the scarce non-trivial theorems on learning nets, it is important that serious side-conditions be made explicit. The language of finite automata is used to mathematicize the problem of adaptation sufficiently to remove some ambiguities of Brindley's approach. We close the paper by relating our framework to other formal studies of adaptation.

This report describes a theory and corresponding model for the neural basis of language. A detailed functional description will be given for the following elementary visual-linguistic processes: (1) the selection and neural encoding of patterns from the visual field; (2) the representation of visual experience in memory; (3) the mechanisms of association between different types of visual and verbal information including (a) naming of visual images, (b) naming of positional relationships between objects, (c) naming of size and shape attributes of objects, and (d) imaging of pictorial information which was previously stored in memory; (4) the neural representation of phrases and simple sentences; (5) the recognition of simple sentences and the concept of meaning; and (b) verballydirected recall of visual experience. Strengths and weaknesses of the model are discussed. Part 1 of this paper contains a complete set of operational definitions. The neural networks aie described, and several alternate control strategies for these networks are considered. Part 2 gives a detailed description of computersimulation studies of the proposed model. Processes demonstrated by the computer simulations are: (1) verbally directed recall of visual experience; (2) understanding of verbal information; (3) aspects of learning and forgetting; (4) the dependence of, recognition and understanding on, contextual information; and (5) elementary concepts of sentence generation. The simulation studies are based on one particular choice of control functions.

The paper presents the fundamentals of the theory of neural nets with loops based on functional matrices, which are a generalization of the state transition matrices for a net. Problems of analysis and synthesis; given a net, find its functional matrix and vice versa, are treated for probabilistic and deterministic nets. Questions about universal nets, oscillations, and stability are studied for deterministic nets. Reduction of probabilistic nets with loops is considered. It is shown that any probabilistic net with loops can be duplicated by a deterministic net with loops plus a probabilistic loop-free encoder. The motivation for the work is a search for formulation of the general theory of neural nets that could be tied to the theory of triadic intensional relations, as suggested by Warren S. McCulloch.

Several modeling techniques have been used to classify the low-back disorders (LBDs) risk associated with the industrial jobs. Many researchers have demonstrated the use of artificial neural networks (ANNs) to predict musculoskeletal disorders risk associated with occupational exposures. In order to improve the accuracy of LBDs risk classification, this paper proposes to use the support vector machines (SVMs), a machine learning algorithm used extensively in the last decade. The results of SVMs based ensemble classification approach to classify the low-back disorders (LBDs) risk associated with the industrial jobs are presented. Four different kernels (i.e. the stump kernel, perceptron kernel, Laplacian kernel and exponential kernel) were used to create infinite ensemble of SVMs and their performance have been compared with the SVMs, AdaBoost SVMs, and a backpropagation neural network. The results suggest an increased performance by stump and Laplacian kernel in comparison to the radial basis function and polynomial kernel based SVMs, AdaBoost SVMs, and the back propagation neural network. Highest classification accuracy of 77.01% was achieved by Laplacian kernel based SVMs in comparison to 71.3% and 74.7% by radial basis function kernel based SVMs and back propagation neural network respectively.

Robust control for nonlinear uncertain systems has been solved for matched uncertainty but has not been completely solved yet for unmatched uncertainty. This paper developed a new method in which an adaptive radial basis function neural network is used to compensate for the effects of unmatched uncertainty in the framework of integral sliding mode control. The stability of the whole system is guaranteed by the Lyapunov method. The adaptation algorithm of the network is also derived by the Lyapunov function so that its convergence is also guaranteed. A numerical example is used to show the effectiveness of the proposed method. Improvement against existing methods is also demonstrated.

This paper discusses possibilities of the realization of some algebraic operation by means of neural nets. The neuron model assumed to be used is modified for this purpose.

This paper reports results of an investigation of the problem of vertebrate color vision by means of a theoretical model, which, although it uses one kind of receptor, can be adapted to a multireceptor concept. It is assumed (1) that the time constant of the change of the receptor potential conveys the color information of the stimulus, whereas the magnitude of the potential is correlated with stimulus intensity and (2) that a group of cells, tentatively identified as ganglion cells, are associated with each receptor field. These cells fire only if the time constant falls within a certain range. Thus, the visual spectrum is divided into regions and the information is transmitted to the central nervous system. Wave length discrimination in the theoretical model is accomplished by one kind of retinal neural nets that are biased differentially. An analog computer was used in this initial phase of the investigation. Care has been taken to ensure that the model satisfies current anatomical and physiological knowledge. It has produced results similar to Granit's (1955) spectral sensitivity and Kelly's (1961) amplitude sensitivity curves. The model, which will predict “subjective color phenomena” at appropriate frequencies, has raised questions amenable to psychophysiological techniques.

McCulloch showed that certain redundancy networks composed of neuron-like elements exhibit the property of "logical stability" even though the individual neurons may undergo transformation of states caused by shift in neural threshold. It was speculated that the stability of neural nets contributes to the improved behaviorial reliability of human beings, a problem that was originally raised by Von Neumann. This paper extends the concept of fallible neurons to conventional logic modules as found in technological systems. It is shown that in a single logic module, transformation of states may occur due to malfunction of the internal components and/or drift in the thresholds of the signal and control inputs. The existence of multiple states in a single module lends itself to the possibility of constructing networks with redundant modules which are functionally more stable than the individual modules. Finally, it is shown that suitably chosen, logically stable structures may be used to replace the individual modules in order to perform logical operations with greater reliability.

This paper states two equivalent conditions from which modularity of the latticeG of decision effectsE can be derived. It may be seen as a supplement to Ludwig's approach [5] to an axiomatic foundation of physical theories. As a consequence of these conditions every filterT E is a self-adjoint projector on the Hilber spaceB′ generated by the decision effects.

The paper presents a new approach to the classification of rolling element bearing faults by implementing Artificial Neural Network. Diagnostics of rolling element bearing faults actually represents the problem of pattern classification and recognition, where the key step is feature extraction from the vibration signal. Characterization of each recorded vibration signal is performed by a combination of signal's time-varying statistical parameters and characteristic rolling element bearing fault frequency components obtained through the frequency spectrum analysis method. The experimental data is collected for four bearings at three different speeds. The sensor is located at three different positions for each bearing. Both time domain and frequency domain signals were measured. Thus the data was three time spectrums and three frequency spectrums for each speed for a bearing. The entire data set comprised of 72 (6 x 3 x 4) data. The time domain signal was comprised of 8192 samples and extracting these features from a huge data set was difficult. To overcome this difficulty the 8192 samples were split into 32 bins each containing 256 samples. Two Network RBFN and PNN are used to classify the bearing defects . The entire process of splitting and evaluating the seven features was coded in MATLAB. From these seven features the most suitable features are for explaining the intensity of the defect is discussed . Key Words: Feature Extraction; Fault Frequencies; Roller Bearing; Bearing fault; Crest Factor; Variant; Radial Basis Function Network (RBFN); Probabilistic Neural Network (PNN) DOI: 10.3329/jme.v40i2.5353 Journal of Mechanical Engineering , Vol. ME 40, No. 2, December 2009 119-130

Part I (P. H. Greene,Bull. Math. Biophysics,24, 247–275, 1962) discussed a number of formal properties of animal behavior, and presented evidence that these properties would follow naturally from a model in which patterns of neural activity in perception or motor action constituted the resonant responses of linear neural networks. Equations were derived for parameters characterizing networks which would possess desired resonant responses. These equations expressed purely mathematical requirements. The present paper shows that a simple neural model would be entirely adequate to meet these requirements. According to this model, an input locus may become functionally connected to a particular resonant response mode by firing at a frequency which comes to approach the resonant frequency of that mode. The information in a complicated “cell assembly” of the type considered could be transmitted through a nerve tract by a very simple frequency code. One neurological guess is that frequency-coded inputs excite the transients in dendritic networks. If the amplitude of the pattern becomes large, as it would near resonance, the all-or-none axonal response would become excited. This axonal response would tend to augment resonant patterns and disrupt other patterns, for a reason inherent in any linear network. Since resonant responses are automatically present in any linear network, unless special processes suppress them, they must have led to overt behavior in animals first possessing such networks. Evolution either suppressed this feature or exploited it. Since its properties resemble those of animal behavior, the latter might be suspected. Some implications are presented regarding what a physiologist might have to look for when he studies a neural system.

There are at present only a few exact solutions of rather simple problems in coupled thermoelasticity. For many problems important from a practical point of view no exact solutions are known. The present paper deals with a convenient method of approximate solutions of these problems. As guideline in the use of this method for the solutions of quasistatic, thermoelastic-coupled problems a numerical example is presented, where the effect of coupling is demonstrated.

An accurate fault distance and direction estimation based on application of artificial neural networks for protection of doubly fed transmission lines is presented in this paper. The proposed method uses the voltage and current available at only the local end of line. This method is adaptive to the variation of fault location, fault inception angle and fault resistance. The Simulation results show that single phase-to-ground faults (both forward and reverse) can be correctly detected and located after one cycle after the inception of fault. Large number of fault simulations using MATLAB® has proved the accuracy and effectiveness of the proposed algorithm. The proposed scheme has significant advantage over more traditional direction relaying algorithms viz. it is suitable for high resistance fault. It has the operating time of less than 1.5 cycles. The proposed scheme allows the protection engineers to increase the reach setting up to 90% of the line length i.e. greater portion of line length can be protected as compared to earlier techniques in which the reach setting is 80-85% only.

In this paper, a new fault diagnosis method using an adaptive neural network for automotive engines is developed. A redial basis function (RBF) network is used as a fault classifier with its widths and weights on-line adapted to cope with model uncertainty and time varying dynamics caused by mechanical wear of engine parts, environment change, etc. Five different sensors are investigated for an automotive engine including throttle angle, manifold pressure, manifold temperature, crankshaft speed and engine torque. The engine data is acquired from a one-litre Volkswagen petrol engine test bed under different operating states, and then simulated multiplicative faults are superimposed. The real data experiments confirm that sensor faults as small as 2% can be detected and isolated clearly. The developed scheme is capable of diagnosing faults in on-line mode and can be directly implemented in an on-board engine diagnosis system.

This paper presents analytical stabiltity analysis of turning using non-linear force-feed model in two dimensions. Most of the existing analytical models ignored the effect of static feed term on the regeneration phenomenon. In practice there is a marked effect of feed on stability due to force variation. The modified analytical equations for cutting insert using three dimensional too geometry are obtained by considering relative motion of tool with respect to a two dimensional elastic model of work-piece. The critical stability limits obtained as a function of feed are confirmed with time domain analysis. Experiments are conducted on a flexible work-piece at varying feed conditions. The measured cutting forces show a marked effect of feed on stability. A neural network models developed to obtain the critical depth of cut at various values of operating speeds and feeds. Key words: Dynamics modeling; Tool geometry; Force-feed; Stab. DOI: 10.3329/bjsir.v45i2.5710Bangladesh J. Sci. Ind. Res. 45(2), 129-132, 2010

Meaningful factor analysis and algebraic operations during stimulation, learning, and discrimination experiments have been performed on averaged evoked potential responses, suggesting that, at least under some circumstances, the signal space of average evoked potentials is linear. Alternatively, the "behavior" of simulated neural nets is defined as the observation of an average over an ensemble of the trajectories of solutions of interconnected nonlinear dynamical systems. This behavior is a mathematical counterpart of the physiological macropotential observations. In this paper, a mathematical model corresponding to the ensemble average over an unconnected set of statistically distributed linear elements suggests duplication of both the simulated neural net and the neurophysiological findings. In contrast with the simulated neural network, the statistical properties of this model are amenable to analysis. The model suggests experiments of the prediction and control of multidiscrimination experiments in cats and provokes questions on the significance of the specification of detail in different levels on the structural hierarchy of the brain.

Many of the observed oscillatory phenomena in biological systems are the result of endogenous neural activity and occur in the absence of any kind of proprioceptive feedback. In certain simple animals such activity is manifested in a stereotyped behavioral response of considerably longer duration than the evoking stimulus. For instance, in Tritonia gilberti (a slug-like nudibranch) even the briefest touch by a predator triggers a powerful escape motion consisting of alternating dorsal and ventral flections of the animal’s body. Physiological studies by A. O. D. Willows revealed that the mechanism responsible for this behaviour is a cooperative effect which requires the interaction between three small pools of neurons. In our earlier publications we have proposed a netlet theory for investigating the dynamics of interacting neuronal ensembles. We found that a variety of cooperative phenomena, such as hysteresis effects, phase transitions, resonances and entrainments exist in such structures. By a combination of mathematical analysis and computer simulation it is shown in this paper that hysteresis effects together with a known property of neurons, namely, accumulating hyperpolarization, give rise to the kind of behavior observed in Tritonia and shed light on some features found in the single unit firing records.

This paper discusses the use of Distance based optimal designs in the design of experiments (DOE) and artificial neural networks (ANN) in optimizing the stacking sequence for simply supported laminated composite plate under uniformly distributed load (UDL) for minimizing the deflections and stresses. A number of finite element analyses have been carried out using Distance-based optimal design, for training and testing of the ANN model. The deflections and stresses were found by analyses which were done by finite element analysis software. The ANN model has been developed using multilayer perceptron (MLP) back propagation algorithm. The adequacy of the developed model is verified by coefficient of determination (R). The sensitivity analysis has been performed to study the behavior of the laminated composite plate. The results obtained from the ANN model are compared with the finite element results. For various fibre orientations, deflections and stresses analyses are performed to get the optimal fibre orientations. A verification tests are also performed to prove the effectiveness of the ANN technique after the optimum levels of fibre orientations are determined. The confirmation experimental results show that deflections and stresses are very good agreed with the finite element (FE) results. Consequently, the Distance-based optimal set of laminates and ANN are shown to be effective for optimization of stacking sequence of laminated composite plates.

This paper describes an investigation of the practicality of describing pictures using linguistic structures other than strings-in particular, using labelled directed graphs (“webs”). A parsing program for a class of “neural network” pictures is demonstrated.

The first portion of this paper is tutorial. Beginning with a standard definition of an abstract pattern-recognition machine, "learning" is given a mathematical meaning and the distinction is made between supervised and unsupervised learning. The bibliography will help the interested reader retrace the history of learning in pattern recognition. The exposition now focuses attention on unsupervised learning. Carefully, it is explained how problems in this subject can be viewed as problems in the identification of finite mixtures, a statistical theory that has achieved some maturity. From this vantage point, it is demonstrated that identification theory implies unsupervised learning is possible in many important cases. The remaining sections present a general method for achieving unsupervised learning. Other authors have proposed schemes having greater computational convenience, but no method previously published is as inclusive as the one revealed here, which we demonstrate to be effective for all the many cases wherein unsupervised learning is known to be possible.

This paper describes two tests for the existence of a linear state assignment for binary input sequential machines. The first test is based on ``transfer sequences'' which map one stationary state of a machine onto another stationary state. It is shown that from a minimal transfer sequence of a machine one can read off directly a linear assignment of this machine if one exists. The second test uses ``ignorance'' computations by means of partitions and previously developed results about sequential machine structure.

Neural firing is herein considered to be a sequence of changes of state of the neuron. Thus, a state diagram analysis of the neuron or of neural nets would seem to be a more natural procedure to follow than would simple logic element analysis. This paper describes a theoretical analysis of the individual neuron, and compares this with both theoretical and experimental analyses of an artificial neuron (a neuromime slightly modified from the Harmon circuit to provide more synaptic actions). The analysis then proceeds to include simple neural networks. The result of such analysis is in the form of a set of state transition tables, from which actual output firing patterns can be obtained. In the experimental phases of this work, input patterns were produced and output was monitored with a linc digital computer. This experimental technique may be developed to greater sophistication for future experiments.

Substantial physiological evidence indicates that neuron thresholds and synaptic weights in living creatures are adjusted by mechanisms quite different from those that have ordinarily been proposed in neural net investigations. This paper presents a theoretical model of the plastic neuron in which threshold and synaptic weights are adjusted solely on the basis of the time history of afferent and efferent activity of the neuron. Physiological, psychological and mathematical evidence is presented which supports the postulate that each neuron in living creatures is an autonomous, dynamically self-adjusting unit which is advised (not directed) by higher centers during the adjustnent process. The model duplicates much of the behavior of neurons in experimental preparations, and simulations of small nets have yielded learning behavior apparently similar in some respects to that of living creatures.

A new perspective on three papers coauthored by Warren McCulloch emphasizes the paradigmatic role to be played in brain theory by action-oriented studies of distributed computation in somatotopically organized neural networks.

The class of systems described by differential equations containing impulses is of interest because most models for biological neural nets belong in that category as well as most pulse frequency modulation systems. In this paper an extension of Liapunov's second method is presented which can be used for the investigation of the stability of such systems. This is obtained by introducing a positive definite function V(x) , which decreases during the occurrence of an impulse and remains constant or decreases during the free motion of the system.

Networks containing neuronal models of the type considered in the previous paper can be described by a set of first order differential equations. Steady-state solutions and the stability of these solutions to small perturbations can be obtained. Networks of physiological interest which give rise to second, third and fourth order linear equations are analysed in detail. Conditions are derived under which such networks can be condensed into a single neuron of similar order. Simple mechanisms for memory storage, for the generation of oscillatory activity and for decision making in neural systems are suggested.

Memory neural networks exhibit promising results for use as adaptive controllers for systems involving nonlinear time-delayed dynamics. By associating a memory neuron to each network neuron, we alleviate the requirement for storing and recurrently feeding a nonlinear plant's past histories in order to adaptively control the system. Past attempts at designing a missile controller with a memory neural network produced encouraging results for the estimation facet of system modeling. This paper presents the design and simulation of a model reference adaptive controller for a missile system using memory neural networks.

This paper presents an efficient method for signal classification from a system of multiple artificial neural networks (ANN) using wavelets. The method performs feature extraction via the wavelet transform of the underlying signal and presents the resulting coefficients to a hybrid neural network for classification. The hybrid network consists of three single neural networks; two of the networks are provided with magnitude and location information of the coefficients, and are trained with self-organizing rules. Their outputs are then presented to the third network for pattern recognition and classification. Experimental results illustrating concept feasibility for acoustic signal classifications are included.

The neural engineering utility with adaptive algorithms (NEUWAA) is a machine-based intelligence system for automatic test equipment, which integrates various technologies in an adaptive fault-detection environment. Computer enhancements and mathematical algorithms allow for the use of man-machine and intelligent applications. The human/machine interface optimizes the test environment by providing state-of-the-art adaptive algorithms to streamline test sequences and diagnostics. NEUWAA employs state-of-the-art diagnostics methodologies coupled with self-organizing evolution to provide an efficient test environment at any level. Highlighted visual images with dialogue of the unit under test provide interactive fault-isolation including guided-probe sequences to streamline fault/diagnosis. The system is completely interoperable with all other standard software packages. This paper outlines the neural engineering utility with adaptive algorithms system including its various characteristics and the techniques involved in its creation. >

The purpose of this paper is to review briefly some of the literature on the use of Procion dyes in Crustacea, to illustrate some of our work on the stomotogastric ganglion of the lobster, and to describe how computer technology can be used to reconstruct and quantitatively analyze dye-filled cells.

Summary For pattern recognition, linear separation networks have been proposed in which the contribution of an input to an association unit is made proportional to the logarithm of the conditional probability of that input given the output of the association unit. Such networks achieve useful discrimination for only the limited condition in which inputs are independent. In practice the output of such a network exceeds the log conditional probability of its corresponding pattern by a large uncalculable quantity; consequently selection cannot be made by comparing the outputs of networks with a calculable significance threshold-one can only select the largest output; a further consequence is that local feedback from output to input does not give the association unit any useful property of autonomous classification. This paper examines a network in which the contribution of an input is made proportional to the Shannon information between that input and the output. Both the above difficulties then disappear, and local feedback causes an association unit to construct a useful class of inputs without the need of an external teacher which ‘knows the answers’. The classification behaviour of such an information network is examined for a number of simple examples in which inputs do not obey an Independence Relation. The results bear some resemblance to recently discovered properties of neurones in the cerebral cortex. The theory suggests that there must be an effect between neurones which has not so far been observed, and a way of looking for it. The theory offers an explanation of some known perceptual phenomena.

A list of important features of animal behavior related primarily to learning has suggested to many investigators that an important aspect of brain function is the establishment or modification of functional connections between neural elements. In this paper a considerably, more inclusive list of behavioral features suggests that another important aspect may be the utilization of patterns of activity constituting the resonant responses of linear networks in the brain. To account for the longer list on the basis of connections requires additional assumptions, while both lists follow immediately from the second mechanism. An input locus may become functionally connected to a particular response mode by firing at a frequency which comes to approach the resonant frequency of that mode. The information in a complicated “cell assembly” of the type considered could be transmitted through a nerve tract by a very simple frequency code. One neurological guess is that frequency-coded inputs excite the transients in dedritic networks. It the amplititude of the pattern becomes large, as it would near, resonance, the all-or-none axonal response would become excited. This axonal response would tend to augment resonant patterns and disrupt other patterns, for a reasonal inherent in any linear network. It is shown how the mechanism might be related to the list of important behavioral features, and a numerical illustration is provided. Since this mechanism is automatically present in any linear network, unless special processes suppress it, it must have led to overt hehavior in any animal, possessing such networks. Evolution either suppressed this feature or exploited it. Since its properties resemble those of animal behavior, the latter might be suspected.

This paper examines algorithms and techniques for use in a complete FLIR target cuer/recognizer. The application is the Air Force Theater Missile Defense Eagle Smart Sensor and Automatic Target Cuer/Recognizer (TESSA) program. The data used for this research are 1st generation FLIR images collected from an F-15E. The database contains thousands of images with various target arrangements. The specific target of interest is a mobile missile launcher, which will be defined as the primary target. The goal is to locate all vehicles (secondary targets) within a scene and identify the missile launchers. The system designed includes an image segmenter, region cluster algorithm, and classifier. Conventional algorithms in conjunction with neural network techniques are used to form a complete ATR system. Some of the conventional techniques include hit/miss filtering, difference of Gaussian filtering, and region clustering. A neural network (multilayer perceptron) is used for classification. These various algorithms are tested and combined into a functional ATR system. Overall target detection rate (cuer) is 84% with a 69% accurate primary target identification (recognizer) rate. Furthermore, the false alarm rate (a non-target cued as a target) is only 2.3 per scene. The research will be completed with a 10 flight test profile using an F-15E and will collect 3/sup rd/ generation FLIR images for use with these algorithms.

A diesel engine fault diagnosis method based on GA-PNN (genetic algorithm-probabilistic neural network) is proposed in this paper, which aims at overcoming the existing shortcomings in traditional diesel engine fault diagnosis methods. Eight kinds of common fault are selected as standard faults identifications, including fuel supplying in different percentage and needle valve sticking, and eight characteristic values are chosen as fault diagnosis indexes, such as maximum pressure. The fault diagnose of diesel engine is conducted by method based on Probabilistic Neural Network and Genetic Algorithms. The testing simulation is completed with data provided by the North China institute of engine. Results show that the fault diagnosis method designed is valid, and compared to traditional methods, the method proposed has certain superiority, which is easy in operation and low in cost.

The excellent information processing in a human brain is considered to depend upon its association mechanisms. To simulate this function, we propose in this paper a model of the neural network named “Associatron” which operates like a human brain in some points. Associatron stores many entities at the same place of its structure, and recalls the whole of any entity from a part of it. From that mechanism some properties are derived, which are expected to be utilized for human-like information processing. After the properties of the model have been analyzed, an Associatron with 180 neurons is simulated by a computer and is applied to simple examples of concept formation and game playing. Hardware realization of an Associatron with 25 neurons and thinking process by the sequence of associations are mentioned, too.

Deep Learning usage is spread across many fields of application. This paper presents details from a selected variety of works published in recent years to illustrate the versatility of the Deep Learning techniques, their potential in current and future research and industry applications as well as their state-of-the-art status in vision tasks, where their efficiency is experimentally proven to near 100% accuracy. The presented applications range from navigation to localization, object recognition and more advanced interactions such as grasping.

Some principles of information theory are utilized in the design of neural nets of the McCulloch-Pitts type. In particular, problems are considered where signals from several neurons must pass through a single one, thus resulting in a “bottleneck” in the flow of information, an abstract model of the corresponding bottleneck from the retina to the optic nerve. The first part of the paper deals with a construction of a McCulloch-Pitts net in which the redundancy in the messages originating in two neurons is utilized so that the messages can be sent over a single neuron with little loss of information. In the second part, messages from a set of neurons are “pumped” into two channel neurons. The optimum connection scheme is computed for this case, i.e, one resulting in a minimum loss of information. Possible biological implications of this approach are indicated.