Showing papers in "Systems and Computers in Japan in 1991"
TL;DR: A computer system is developed for automatic recognition of continuously spoken words by using only visual data, and the pattern of muscle action is used to recognize the spoken words.
Abstract: While the acoustic signal is the primary cue in human speech recognition, the visual cue is also very useful, especially when the acoustic signal is distorted. A computer system is developed for automatic recognition of continuously spoken words by using only visual data. The velocity of lip motions may be measured from optical flow data which allows muscle action to be estimated. Pauses in muscle action result in zero velocity of the flow and are used to locate word boundaries. The pattern of muscle action is used to recognize the spoken words.
182 citations
TL;DR: This paper evaluates the CINON system and compares it with the Lamport authentication system, an earlier system which uses a one-way data transformation, and finds that CINon's execution speed is faster by a factor of perhaps several hundred to a thousand.
Abstract: A new password-based authentication system CINON is proposed for use in communications and computer systems It employs a one-way function to perform the required authentication of communicating users CINON maintains its security in spite of a wiretap or the theft of a password file, and it is not necessary to replace the correspondents' public passwords CINON can be realized with only a few computations
This paper evaluates the CINON system and compares it with the Lamport authentication system, an earlier system which uses a one-way data transformation In comparison with the Lamport system, CINON's execution speed is faster by a factor of perhaps several hundred to a thousand
50 citations
TL;DR: A new multiscale segment matching method which is applicable to heavily deformed planar shapes and makes it possible to obtain intuitively relevant correspondences even if the shapes have some local heavy deformation.
Abstract: This paper proposes a new multiscale segment matching method which is applicable to heavily deformed planar shapes. First, multiscale representations are obtained using curvature scale space filtering. Then inflection point correspondence is developed between consecutive smoothed shapes. The representation in this paper, unlike the well-known curvature scale space image description, ensures that it always satisfies the consistency of hierarchical segment replacement. Moreover, it requires less processing time and memory allocation. Finally, optimum scale segments are matched by a new multiscale segment matching method proposed herein. In this method, the matching problem is formulated as a minimization problem of the total amount of segment dissimilarity. The minimization problem is solved effectively using dynamic programming. The proposed matching method makes it possible to obtain intuitively relevant correspondences even if the shapes have some local heavy deformation.
45 citations
32 citations
TL;DR: This paper presents the development of a display method to preserve the continuity of organ shapes through the extraction of fuzzy shapes and their shading and shows the effectiveness of the proposed 3-D display method using head MRI images.
Abstract: A three-dimensional (3-D) display of medical images allows better perception of the continuity of structures and spatial relationships than a display of individual slices. These 3-D display methods consist of a contour extraction, a shape reconstruction and a shading. However, the boundary judgment of soft tissues, like a tumor, is difficult even for expert doctors. Especially, 3-D display depends largely on a result of the contour extraction. The 3-D display method is required to display the presentation of the continuity of organ shapes.
This paper presents the development of a display method to preserve the continuity of organ shapes. The display method consists of the extraction of fuzzy shapes and their shading. The extraction of fuzzy shapes is executed by operation of the fuzzy c-means clustering algorithm on medical images and the use of a connectivity of an organ shape. The shading of fuzzy shapes uses a volume rendering, which is analyzed numerically and improved to realize a powerfully interactive 3-D display. The effectiveness of the proposed 3-D display method is shown using head MRI images.
22 citations
TL;DR: Fuzzy learning vector quantization had higher ability in an odor discrimination from the known one than is the case in conventional networks and that an unknown odor was discriminated by FLVQ.
Abstract: Because of its high performance, the learning vector quantization (LVQ) proposed by Kohonen is noteworthy as a method for realizing a neural network. We propose herein a new method of LVQ using fuzzy theory and call it “fuzzy learning vector quantization” (FLVQ).
FLVQ algorithm is as simple as that of LVQ, and its capability of pattern recognition is higher than that of a conventional neural network. Although it is difficult for conventional neural networks to discriminate an input pattern of an unknown category from those of known ones, FLVQ can do it. Since reference vectors of FLVQ are described by use of fuzzy sets and their membership functions are obtained from learning, the data features can be effectively extracted using FLVQ.
We have used FLVQ in an odor pattern recognition system and compared its capability with those of conventional neural networks. As a result, it was confirmed that FLVQ had higher ability in an odor discrimination from the known one than is the case in conventional networks and that an unknown odor was discriminated by FLVQ.
22 citations
TL;DR: In this article, a new parallel-hierarchical neural network model was proposed to enable motor learning for simultaneous control of both trajectory and force, by integrating Hogan's control method and the author's previous neural network control model using a feedback-error-learning scheme.
Abstract: This paper proposes a new parallel-hierarchical neural network model to enable motor learning for simultaneous control of both trajectory and force, by integrating Hogan's control method and the author's previous neural network control model using a feedback-error-learning scheme. Furthermore, two hierarchical control laws which apply to the model are derived by using the Moore-Penrose pseudo-inverse matrix: one is related to the minimum muscle-tension-change trajectory; and the other is related to the minimum motor-command-change trajectory. The human arm is redundant at the dynamics level since joint torque is generated by agonist and antagonist muscles. Therefore, acquisition of the inverse model is an ill-posed problem. However, the combination of these control laws and feedback-error-learning resolves the ill-posed problem. Finally, the efficiency of the parallel-hierarchical neural network model is shown by learning experiments using an artificial muscle arm and computer simulations.
19 citations
TL;DR: A method to synthesize facial motion images based on given information on text and emotion is investigated, which utilizes the analysis-synthesis image coding method and synthesizes reasonably natural facial images.
Abstract: A facial motion image synthesis method for intelligent man-machine interface is examined. Here, the intelligent man-machine interface is a kind of friendly man-machine interface with voices and pictures in which human faces appear on a screen and answer questions, compared to the currently existing user interfaces which primarily uses letters. Thus what appears on the screen is human faces, and if speech mannerisms and facial expressions are natural, then the interactions with the machine are similar to those with actual human beings.
To implement such an intelligent man-machine interface it is necessary to synthesize natural facial expressions on the screen. This paper investigates a method to synthesize facial motion images based on given information on text and emotion. The proposed method utilizes the analysis-synthesis image coding method. It constructs facial images by assigning intensity data to the parameters of a 3-dimensional (3-D) model matching the person in question. Moreover, it synthesizes facial expressions by modifying the 3-D model according to the predetermined set of rules based on the input phonemes and emotion, and also synthesizes reasonably natural facial images.
16 citations
TL;DR: This paper proposes a new method of locating an edge in a digital image to sub-pixel precision by using gray-level information in the neighborhood of the edge alone using the least-squares method.
Abstract: This paper proposes a new method of locating an edge in a digital image to sub-pixel precision by using gray-level information in the neighborhood of the edge alone. The location of the edge is determined by fitting the first derivatives of gray-level values to a normal-distribution curve using the least-squares method and by regarding the location of the extreme value of the curve as the location of the edge. By comparing the proposed method with five conventional methods it is shown that the method requires the smallest amount of computation generally and minimum systematic errors in a sharp edge. The errors in an edge location measurement caused by image noises are quantized and are shown by a graph so that errors under an arbitrary condition can easily be estimated.
16 citations
TL;DR: It is demonstrated that the proposed system has a fairly high detection performance for the cancer lesion with fold convergence, and a new feature parameter called convergence index is defined to represent quantitatively the extent of convergence of the folds on the stomach wall.
Abstract: This paper describes an automatic detection system which extracts pathological areas from double contrast stomach x-ray images developed for computer-assisted diagnosis of stomach x-ray images. This paper examines the image for stomach cancer, especially the convergent fold pattern of the stomach wall which is inherent to the early stage of cancer, and extracts the pattern from the x-ray image.
The actual processing procedure is divided largely into the following four stages: (1) extraction of the border of the fold region; (2) extraction of contrasted regions of the bowel; (3) evaluation of the fold convergence; and (4) extraction of the lesion candidates.
In this paper, a new feature parameter called convergence index is defined to represent quantitatively the extent of convergence of the folds on the stomach wall. The detailed computation procedures for the convergence index are presented. Finally, an experiment is presented for the actual 25 examples of double contrast stomach x-ray images, and it is demonstrated that the proposed system has a fairly high detection performance for the cancer lesion with fold convergence. No study has been presented on the recognition of the pathological pattern for the stomach cancer in the double contrast stomach x-ray image, and the result in this paper is expected to provide an important clue to the realization of the future automatic diagnosis of the stomach x-ray images.
14 citations
TL;DR: This paper considers the complete network with the sense of direction where at most fp processors are in the fail-stop condition and presents an algorithm which solves the leader election problem with the communication complexity O(n + k) (k is the number of initiating processors).
Abstract: Consider a directed Hamilton cycle H on a complete network. If each processor can distinguish its incident links by the distance on H, the Hamilton cycle is said to have the sense of direction. It is known that the communication complexity of the leader election problem can be reduced by using the sense of direction when there exists no failure in the network.
This paper considers the complete network with the sense of direction where at most fp processors (fp < n/2, where n is the total number of processors in the network) are in the fail-stop condition and presents an algorithm which solves the leader election problem with the communication complexity O(n + k.fp) (k is the number of initiating processors). The result implies that the sense of direction can be used to reduce the communication complexity of the leader election problem also in the complete network with fail-stop processors. It is shown also that the communication complexity of the presented algorithm is optimal within a constant factor.
TL;DR: This paper presents efficient distributed algorithms on an asynchronous network for the following problems: finding bi-connected components, finding cutpoints, finding bridges, testing for bi- connectedness and finding strongly connected components of a directed graph defined on a network.
Abstract: This paper presents efficient distributed algorithms on an asynchronous network for the following problems: finding bi-connected components, finding cutpoints, finding bridges, testing for bi-connectedness and finding strongly connected components of a directed graph defined on a network. All these distributed algorithms use a depth-first search tree having an arbitrary processor in the network as its root. The communication complexity of these algorithms is O(nlogn+e) and their ideal-time complexity is O(nG(n)), where n and e represent the numbers of processors and links, respectively, and G(n) is almost a constant. It is shown also that a lower bound for the communication complexity of the five problems is O(e) and a lower bound for their ideal-time complexity is O(n).
TL;DR: In this paper, the authors define characteristic parameters for a multilayer feedforward neural network to examine its information transmission capabilities and, using those parameters, clarifies the behavior of and problems with the conventional method using random initial values.
Abstract: There is no known systematic method for setting the initial values of weights and thresholds for backpropagation learning in multilayer feedforward neural networks. Thus, the trial-and-error method of providing random numbers distributed evenly over a certain range has come to be commonly used.
This paper defines characteristic parameters for a multilayer feedforward neural network to examine its information transmission capabilities and, using those parameters, clarifies the behavior of and problems with the conventional method using random initial values. One such problem is that the active range of a unit narrows rapidly as the number of inputs is increased. In view of this, a new method is proposed for determining initial values which expand the active range; and its usefulness is shown through numerical experiments.
TL;DR: The method proposed efficiently diagnoses multiple stuck-at faults in combinational circuits by adding preprocessing to further reduce the number of probes and postprocessing to improve the diagnosis resolution.
Abstract: To reduce VLSI turnaround time and cost, an efficient fault diagnosis method must be developed. Electron-beam probing, which allows observation of the signal values within a chip, is one of the promising approaches for fault diagnosis. However, it is strongly desired that the number of probes be reduced; otherwise, a great deal of effort results.
The electron-beam probing, guided by deduction based on implications of internal signal values [6], is considered one of the effective solutions to this problem. Here this idea is built upon by adding preprocessing to further reduce the number of probes and postprocessing to improve the diagnosis resolution. The method proposed efficiently diagnoses multiple stuck-at faults in combinational circuits. Using computational experiments it is also shown that almost all of the faults are diagnosed by probing about 15 percent of the nets when test patterns for detecting single stuck-at faults are used.
TL;DR: It is shown that complete depth amplitude reproduction of stereoscopic images through the use of existing CRT display systems is difficult, and a method for overcoming discontinuous depth reproduction is proposed and its effectiveness is confirmed.
Abstract: Three-dimensional visual communications, realized with stereoscopic images, can achieve total display realism. To create a three-dimensional (3-D) transmission and display system, the large amount of information typically contained in natural images must be reduced and coded by an efficient technique appropriate for stereoscopic images.
As the initial step in realizing such a coding scheme, statistical characteristics of stereoscopic images and image characteristics viewed from the aspect of human stereopsis are presented.
First, the possibility of data compression that exploits the cross correlation between right and left images is shown. Based on this, a method that shifts one image horizontally and subtracts it from the corresponding area of the other is proposed and evaluated. Next, it is shown that complete depth amplitude reproduction of stereoscopic images through the use of existing CRT display systems is difficult. This is due to the poor horizontal resolution of the forementioned systems even though the depth spatial frequency bandwidth of stereoscopic images can be reduced without deteriorating depth shape reproduction accuracy. Finally, a method for overcoming discontinuous depth reproduction is proposed and its effectiveness is confirmed.
TL;DR: This paper proposes a parallel execution mechanism with the refinement between machine instruction level to the register transfer level and the basic configuration of the computer is described for the prototype model based on three classes of machine instructions.
Abstract: This paper proposes a parallel execution mechanism with the refinement between machine instruction level to the register transfer level. In the proposed mechanism, the processor is function-partitioned into a number of processing units, and instruction streams are assigned independently to the processing units. According to the arc indicating the dependency in the control flow graph, the processing units exchange the asynchronous control signals at a high speed. Compared with the synchronous parallel execution mechanism such as VLIW, the proposed mechanism can extract the parallelism in a flexible and detailed way.
First, the general parallel computation model based on the proposed mechanism is described. Then the basic configuration of the computer is described for the prototype model based on three classes of machine instructions. A simple evaluation is presented through execution examples of sample programs.
TL;DR: An efficient verification method based on the single-place zero-reachability problem of time Petri nets that generates reduced state space sufficient to check the possibility of the firings of all enabled transitions and compares the performance of both the proposed method and the previous method.
Abstract: A verification method based on time Petri nets automatically proves the correctness of quantitative timing properties for hard real-time systems. However, the previously proposed method sometimes takes a very long time or requires considerable memory space to complete verifications. This is because the method generates the whole state space of time Petri nets.
This paper proposes an efficient verification method based on the single-place zero-reachability problem of time Petri nets, but it may be restricted by given time Petri nets. The single-place zero-reachability problem can be solved only by checking whether or not some prescribed transitions can finally fire. This does not require extracting the whole state space of time Petri nets.
The method proposed herein: (1) reduces the verification problem to the single-place zero-reachability problem of time Petri nets; and (2) generates reduced state space sufficient to check the possibility of the firings of all enabled transitions. This paper also contains an example of verifying a simple LAN protocol and compares the performance of both the proposed method and the previous method.
TL;DR: This identification method uses a facial three-dimensional shape that is extracted from an absolute range map that condenses the whole feature included in a surface into a small number of points (vertices), and features can be used efficiently for the identification.
Abstract: This paper proposes a new method for automatic identification of human faces. This identification method uses a facial three-dimensional (3-D) shape that is extracted from an absolute range map. The important issue is to determine what to use as feature vector components for the identification of 3-D objects. There have been several approaches to identify 3-D objects, but in the identification of objects that have smooth and complex shape (such as a human face) such approaches did not use the whole feature included in an object, or a long time was required for the identification because of the large number of feature vector components needed.
To solve these problems, an approximation of a facial surface with a B-spline surface was made by least-square fitting, and the vertices of the B-spline surface were used as feature vector components for the identification. Since this process condenses the whole feature included in a surface into a small number of points (vertices), features can be used efficiently for the identification. By this method, identification accuracy of 98.8 percent is obtained for 33 persons (5 data/person).
TL;DR: The algorithms developed for three-dimensional bone section removal and designed for high speed and ease of operation are described and an actual plastic surgery case to which the algorithms were applied is presented.
Abstract: With the current rapid advances in computer graphics and CAD/CAM technology, considerable research is being done on the display of three-dimensional images, and this includes the medical field. One typical medical application is simulated surgery, of which bone cutting is an essential part. Facilities for removing bone sections two-and also three-dimensionally are required.
This paper describes the algorithms developed for three-dimensional bone section removal and designed for high speed and ease of operation. An actual plastic surgery case to which our algorithms were applied is presented. The three-dimensional images were constructed from multislice CT data, and the simulations were performed using a high-speed graphics processor.
TL;DR: This paper analyzes the typical two crack patterns produced in different materials, one of which is only slightly sticky and the other is very sticky, and presents the behavioral model, having one parameter of stickiness, which is effective for visual simulation of crack patterns.
Abstract: Cracks produced by drying mud-like materials exhibit various patterns of form (including the crack patterns that occur on plate glass and on the ground) according to the stickiness of the materials. It is very interesting and important to develop a behavioral model of cracks, which produces realistic crack patterns, from not only a theoretical but a practical view, e.g., realistic image synthesis in CG.
In this paper, we first analyze the typical two crack patterns produced in different materials, one of which is only slightly sticky and the other is very sticky. We then show some basic assumptions useful for constructing a behavioral model of cracks.
Next, we present the behavioral model having one parameter of stickiness; and presenting several simulated crack patterns, we show that the constructed model is effective for visual simulation of crack patterns.
We then present CG images, synthesized by employing texture mapping techniques, of marble objects, china teacups, and a china vase, all of which have many cracks on their surfaces. Finally, we discuss further improvements for extending the model to one which works on any curved surface and produces more realistic crack patterns according to the shapes of the surfaces.
TL;DR: This paper shows experimentally that a three-layer neural network learns to have a similar capability to that of the Fourier transform method by means of the back propagation learning, during which a set of random dot patterns are repeatedly presented with various positions.
Abstract: In many pattern recognition tasks (e.g., handwritten letters), it is essential to recognize a given pattern invariantly with translation, rotation, and other spatial transformations to which the pattern might be subjected. The Fourier transform method is a well-known method for translation invariant pattern recognition in which the Fourier spectrum of an image is used as an invariant feature.
This paper shows experimentally that a three-layer neural network learns to have a similar capability to that of the Fourier transform method by means of the back propagation learning, during which a set of random dot patterns are repeatedly presented with various positions. Specifically, each unit in the middle layer of the resultant network detects a particular frequency component contained in the given image; and, using the information, the output layer generates an output pattern which is invariant to translation of the image.
TL;DR: A new algorithm can make it possible to reconstruct the surface shape from the information of only the azimuth angles of lighting directions, which is similar to the conventional photometric stereo.
Abstract: As a photometric stereo method, this paper proposes a new algorithm for reconstructing a surface shape under the condition that the information of zenith angles of three light source directions is unknown.
The object has a uniform Lambertian surface which has whole gradients. An illuminating equation consisting of the parameters of light source directions is derived from the information of three shading images. The simultaneous equations can be held at multiple surface elements and be solved by the method of least squares. This method makes it possible to extract the information of zenith angles of three lighting directions and optimized surface gradient distribution of the object.
When the method of least squares is applied, a method to set the initial values for three zenith angles of lighting directions is described, and a method to select surface elements which construct the objective function is proposed. The effectiveness and the validity are confirmed by computer experiments.
While the conventional photometric stereo treats the information of zenith angles and azimuth angles of three lighting directions as known constants, this algorithm can make it possible to reconstruct the surface shape from the information of only the azimuth angles of lighting directions.
TL;DR: The results show that appropriate thresholds can be obtained even for difficult images, and a method to calculate the mean adjacent-pixel number at a high speed by combining the rank filters and histograms so that the amount of computation becomes independent of the number of thresholds.
Abstract: This paper proposes a method of selecting an appropriate adaptive threshold in binarization (0 or 1) of a gray-level image. A measure representing compactness of a connected component in an image, “the mean adjacent-pixel number,” is introduced. The best threshold is determined by taking the maximal of this measure. The method is applied successfully to actual gray-level images. The results show that appropriate thresholds can be obtained even for difficult images (e.g., a small object in a noisy and low-contrast image) which are less successful in conventional adaptive methods (e.g., the gray-level histogram method). Conventional methods using an adaptive threshold generally have a shortcoming in that the amount of computation is proportional to the number of thresholds. This paper also proposes a method to calculate the mean adjacent-pixel number at a high speed by combining the rank filters and histograms so that the amount of computation becomes independent of the number of thresholds.
TL;DR: Examples in which this method is applied to extract title characters from printed pages and to extract a specific part from a map image are described, and it is shown that this method using higher-order features possess sufficient discrimination ability for real texture images.
Abstract: In texture discrimination the basic image processes are considered to choose effective resolution, to determine representative features and to unify subregions having the same representative features. To realize these processes, this paper presents an improved method for discriminating higher-order features by using a self-organized multiresolutional filter.
First, we make a local transformation function with the maximum discrimination ability by a limited-size neighborhood operation. Next, extending the method to the multiresolutional case, we let neighbor pixels of resolution which is optimal for discrimination be chosen automatically. Moreover, we realize unification of texture regions by connecting features represented by these pixel values and by obtaining macro-features.
In this paper we also describe examples in which this method is applied to extract title characters from printed pages and to extract a specific part from a map image, and show that this method using higher-order features possess sufficient discrimination ability for real texture images.
TL;DR: An algorithm is presented which can determine, from the set of faults, the fault for the reference pattern of the supply current, which can be demonstrated experimentally for the TTL logic circuit.
Abstract: This paper considers a method of detecting a fault in the TTL combinational circuit based on the supply current flowing into the circuit. The authors have proposed a fault detection method for the combinational circuit based on the supply current and indicated the feasibility of fault detection. However, the method proposed in the past is based only on the supply current waveform, and it is required to determine the feature vectors for all the measured supply currents. Consequently, to apply the method to the practical fault detection of logic circuits, the computational complexity as well as the computation time in determining the feature vector should be reduced.
From such a viewpoint, this paper proposes the following method of fault detection. Whether or not a fault exists is determined first by the average, the maximum and the minimum of the supply current. Only for the supply current waveform, which cannot be decided as the nonfaulty circuit, is the feature vector employed to make a further decision using the pattern recognition technique. The effectiveness of the method is demonstrated experimentally for the TTL logic circuit. Also, an algorithm is presented which can determine, from the set of faults, the fault for the reference pattern of the supply current.
TL;DR: This paper proposes a model for local motion detection based on the model by Watson et al., integrating the information from a number of spatial frequency channels, and determines the motion velocity from the intersection of straight lines on the velocity plane.
Abstract: Numerous studies have recently been made on the mechanism of human motion perception, especially on the model for the local motion detection using the spatio-temporal filter. The feature of such a model is that it is not required to solve the matching problem which is inherent to the feature matching problem. Another point is that the model can account for several properties of human motion perception. However, earlier models are limited in their performance since the motion is detected using only a single spatial frequency channel. The usual scene contains a large number of spatial frequency components. To detect the motion with a high accuracy, it is desirable to integrate the information obtained from a number of channels. It is suggested by several psychological experiments that an interaction occurs between channels of the human visual system.
From such a viewpoint, this paper proposes a model for local motion detection based on the model by Watson et al., integrating the information from a number of spatial frequency channels. The model determines the motion velocity from the intersection of straight lines on the velocity plane, where the straight line is calculated from the spatio-temporal filter output, to represent the candidate for the velocity. The interaction between channels is represented by drawing the straight lines obtained from various channels on the same velocity plane. The effectiveness of the proposed method is shown by a simulation using a random dot pattern.
TL;DR: In this paper, a multiangled parallelism matching method was proposed for extracting higher-order features composed of relation among lines, where the model of a shape is given as a "direction" at each evaluation point and "displacement" from the previous evaluation point to the evaluation point.
Abstract: To extract geometric features by parallel operation, a directional erosion/dilation operation method by unifying directional feature field and directional parallel operation was proposed and its effectiveness was verified by application to the extraction of buildings, roads, hatching regions, and character regions from a topographical map.
This paper extends this method further and proposes the “multiangled parallelism matching method” for extracting higher-order features (models of shapes) composed of relation among lines. The input is represented by “directional planes.” The model of a shape is given as a “direction” at each “evaluation point” and “displacement” from the previous evaluation point to the evaluation point. Matching is performed by repeating parallel operation for the entire plane of “translation” of a “counter plane” storing evaluated quantity and “addition” of the plant of input direction into the counter plane.
The MAP matching method has a feature that the result of exhaustive matching for directional elastic templates is obtained as consecutive evaluated quantities by parallel operation. The exhaustive template matching, Hough transform, and DP matching method are compared, and the results of recognition experiments of echocardiograms of the heart are shown.
TL;DR: A neural network model which memorizes temporal patterns such as movement patterns of animals and shows that “direct BP learning,” which considers only the direct effect of the output of the hidden unit on the output unit, has almost the same learning performance with less computation time than “simultaneous BP learning.”
Abstract: This paper proposes a neural network model which memorizes temporal patterns such as movement patterns of animals. A transient or periodic signal waveform is memorized and regenerated as a transient response waveform or a waveform of an autonomous oscillation in the continuous-time neural network with recurrent connections. The back-propagation (BP) learning rule, which has been used primarily in the discrete-time neuron model with feed-forward connections, is extended in two ways so that it is applicable to the continuous-time neural model with recurrent connections.
As a result of computer simulation, it is shown that “direct BP learning,” which considers only the direct effect of the output of the hidden unit on the output unit, has almost the same learning performance with less computation time than “simultaneous BP learning,” which is based more strictly on the learning by the steepest descent.
TL;DR: This paper discusses the alphanumeric character recognition using two kinds of PDF (parallel distributed processing) models and shows that FPM with Kullback divergence evaluation can realize a more accurate recognition with fever number of trainings, compared to the conventional BP model.
Abstract: This paper discusses the alphanumeric character recognition using two kinds of PDF (parallel distributed processing) models. The models considered are the FPM (fuzzy partition model) of a new type with multiple input/output units, and the conventional BP (back-propagation) model. Two BP algorithms with different error evaluation functions are applied to those models, and the number of trainings and the recognition rate in the multifont recognition are compared.
As a result of experiment, it is shown that FPM with Kullback divergence evaluation can realize a more accurate recognition with fever number of trainings, compared to the conventional BP model. The reason for the reduction of the number of trainings is discussed when the mutual inhibition in FPM unit as well as the Kullback divergence as the error evaluation function are employed.