Hirokazu Yokoi

Evaluation of Prediction Capability of Non-recursion Type 2nd-order Volterra Neuron Network for Electrocardiogram

The prediction accuracy of QRS wave that show electric excitement by the ventricle of the heart is low in linear predictions of electrocardiogram (ECG) used a conventional linear autoregressive model, and it is a problem that the prediction accuracy is not improved even if the prediction order is set second and third or more. The causes are that QRS wave generated by the nonlinear generation mechanism and the nonlinear components which the linear models cannot predict is included in ECG. Then, Non-recursion type 1st-order Volterra neuron network (N1VNN) and Non-recursion type 2nd-order Volterra neuron network (N2VNN) were evaluated about nonlinear prediction accuracies for ECG. The results of comparing nonlinear predictions of both networks showed that N2VNN is 17.6 % smaller about the minimum root mean square error indicating prediction accuracy than N1VNN.

Improvement of a neural network-based motion generator with bimanual coordination for upper limb prosthesis

We propose to redesign a neural network used as a motion generator with bimanual coordination for upper limb prosthesis in order to improve its learning capability. We assumed that the wearer of the prosthesis was a unilateral amputee. In our previous work, we proposed a prosthesis control system using a neural network that learned bimanual coordination in order to implement smooth motion with both hands. However, the previously proposed system has the problem that a neural network cannot generate the desired motion of the prosthesis in special cases. The reason is that the motion generator calculates the desired posture of the prosthesis from the current posture of the healthy arm only, regardless of the current posture of the prosthesis. We propose to use the current posture of both the healthy arm and the prosthesis as neural network inputs in order to solve this problem. In this article, we show that a single neural network whose input was the current posture of both arms could learn the relationships of the coordinated motions of holding boxes of different sizes, and the newly proposed system can calculate the desired motion of the prosthesis in special cases through computer simulations.

A basic study on a learning motor vehicle using basic elements for neural computer, continuous-time Folthrets

Abstract A neural computer is helpful to automation of motor vehicle driving. In the present paper, a motor vehicle which can be automatically driven by a learning neural computer is called a learning motor vehicle, and as a basic element for the neural computer, a continuous-time Folthret is newly proposed. It was embodied as an electronic circuit, and two of the continuous-time Folthret-embodying electronic circuits were used to constitute a simple neural computer which was mounted on an electric motor vehicle, to make a learning motor vehicle. Finally, to automate regular driving operation, the manufactured learning motor vehicle was used for a driving experiment, in which the neural computer was made to learn the relation between the steering operation of a driver and the surrounding scenes, for examining its autonomous driving performance in reference to the traveling course. As a result, it was confirmed that the computer could learn the steering operation of the driver after about ten trials even though the conditions were limited to constant driving speed, simple scenes, relatively narrow room, static environment, less input information and a learning motor vehicle very simple in neural computer structure, and it was demonstrated that a neural computer was effective to some extent for automation of motor vehicle driving.

A fundamental element for neural computer - Folthret

Abstract This paper presents Folthret as a fundamental element for the neural computer. This element is constructed based on a discrete-time learning threshold element model of the neuron, utilizing a signal called a Fourier series signal onto whose Fourier coefficients information is loaded. To this element, Hebbian learning, membrane potential learning, correlation learning, error-correlation learning, and orthogonal learning can be aplied. Folthret has the following strong points: 1. (1) Just two lines are sufficient for input, 2. (2) calculation of the sum of the weighted input signals and change of the connection weights can be made very easily, and 3. (3) no matter how complicated the network is, it can be realized with simple wiring. Concurrent with this, a change of the network can be made without any difficulty. Accordingly, Folthret is highly advantageous in allowing a neural computer to be implemented electronically.

Evaluation of an optimal design method for a multilayer perceptron by using the design of experiments

We evaluated the performance of an optimal design method for a multilayer perceptron (MLP) by using the design of experiments (DOE). In our previous work, we proposed an optimal design method for MLPs in order to determine the optimal values of such parameters as the number of neurons in the hidden layers and the learning rates. In this article, we evaluate the performance of the proposed design method through a comparison with a genetic algorithm (GA)-based design method. We target an optimal design of MLPs with six layers. We also evaluate the proposed designed method in terms of calculating the amount of optimization. Through the above-mentioned evaluation and analysis, we aim at improving the proposed design method in order to obtain an optimal MLP with less effort.

A neural network based hierarchical motor schema of a multi-finger hand and its motion diversity

This paper presents a neural network based hierarchical motor schema of a multi finger hand to generate suitable behavior for an unknown situation without retraining all neural networks and investigates its motion diversity by changing its input signals. Conventional neural networks are hard to generate desired movements in an unknown situation. Our hierarchical motor schema consists of the two layers. A lower schema is implemented by a recurrent neural network trained with primitive movement patterns and generates a finger movement from a command code sent from the upper schema. The upper schema generates command codes to each finger from a behavior command code such as grasping. We showed that though the lower schemata were fixed, diversity of generated finger movements can be obtained by changing a behavior code of the upper schema through computer simulation.

Perturbational Neural Networks for Incremental Learning in Virtual Learning System

This paper presents a new type of neural networks, a perturbational neural network to realize incremental learning in autonomous humanoid robots. In our previous work, a virtual learning system has been provided to realize exploring plausible behavior in a robots brain. Neural networks can generate plausible behavior in unknown environment without time-consuming exploring. Although an autonomous robot should grow step by step, conventional neural networks forget prior learning by training with new dataset. Proposed neural networks features adding output in sub neural network to weights and thresholds in main neural network. Incremental learning and high generalization capability are realized by slightly changing a mapping of the main neural network. We showed that the proposed neural networks realize incremental learning without forgetting through numerical experiments with a two-dimensional stair-climbing bipedal robot.

A new element for parallel distributed processing network—learning distance product element

One of the factors which affect the convergence velocity in the learning by a parallel distributed processing network is the input/output characteristic of its elements. This paper proposes anew a learning distance product element to substitute for the conventional learning threshold elements, and performs the computer simulation of a three-layered network with two inputs and one output using the learning distance product elements. As a result, it is confirmed that on average, learning is accelerated compared to the use of the learning threshold elements.