Showing papers by "Nobuyuki Matsui published in 2004"

Quaternion neural network with geometrical operators

Abstract: Quaternion neural networks are models in which computations of the neurons are based on quaternions, the four-dimensional equivalents of imaginary numbers. This paper shows by experiments that the quaternion-version of the Back Propagation (BP) algorithm achieves correct geometrical transformations in three-dimensional space, as well as in color space for an image compression problem, whereas real-valued BP algorithms fail. The quaternion neural network also performs superior in terms of convergence speed to a real-valued neural network with respect to the 3-bit parity check problem, as simulations show.

Fault-tolerance in nanocomputers: a cellular array approach

Abstract: Asynchronous cellular arrays have gained attention as promising architectures for nanocomputers, because of their lack of a clock, which facilitates low power designs, and their regular structure, which potentially allows manufacturing techniques based on molecular self-organization. With the increase in integration density comes a decrease in the reliability of the components from which computers are built, and implementations based on cellular arrays are no exception to this. This paper advances asynchronous cellular arrays that are tolerant to transient errors in up to one third of the information stored by its cells. The cellular arrays require six rules to describe the interactions between the cells, implying less complexity of the cells as compared to a previously proposed (nonfault-tolerant) asynchronous cellular array that employs nine rules.

A multilayered feed-forward network based on qubit neuron model

Abstract: With the development of our highly information-oriented society, there is an increasing demand for large-scale and high-level information processing. Toward this goal, studies have sought to create a new computation principle having an information processing ability exceeding the existing Neumann-type computer, such as the creation of a new computation theory or the integration of the frameworks of existing computation theories. As one such approach, quantum neural computing is considered to be interesting, which integrates neural computing and quantum computation. This paper constructs the feed-forward neural network, which is widely used in practice, based on the qubit neuron model. The 4-bit parity-check problem and the general function identification problem are considered. The performance is compared to the feed-forward network based on the conventional neuron model, and it is shown that the proposed model has a higher performance than the conventional model, using the learning diagram composed of convergence rate and the number of learning iterations. The reason for the better performance is also discussed. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(13): 43–51, 2004; Published online in Wiley InterScience (). DOI 10.1002sscj.10342

Effects of Chaotic Exploration on Reinforcement Maze Learning

Abstract: In reinforcement learning, it is necessary to introduce a process of trial and error called an exploration. As a generator for exploration, it seems to be familiar to use the uniform pseudorandom number generator. However, it is known that chaotic source also provides a random-like sequence as like as stochastic source. In this research, we propose an application of the random-like feature of deterministic chaos for a generator of the exploration. As a result, we find that the deterministic chaotic generator for the exploration based on the logistic map gives better performances than the stochastic random exploration generator in a nonstationary shortcut maze problem. In order to understand why the exploration generator based on the logistic map shows the better result, we investigate the learning structures obtained from the two exploration generators.

Fault-tolerant nanocomputers based on asynchronous cellular automata

Abstract: Cellular Automata (CA) are a promising architecture for computers with nanometer-scale sized components, because their regular structure potentially allows chemical manufacturing techniques based on self-organization. With the increase in integration density, however, comes a decrease in the reliability of the components from which such computers will be built. This paper employs BCH error-correcting codes to construct CA with improved reliability. We construct an asynchronous CA of which a quarter of the (ternary) bits storing a cell's state information may be corrupted without affecting the CA's operations, provided errors are evenly distributed over a cell's bits (no burst errors allowed). Under the same condition, the corruption of half of a cell's bits can be detected.

Learning based on fault injection and weight restriction for fault-tolerant Hopfield neural networks

Abstract: Hopfield neural networks tolerating weight faults are presented. The weight restriction and fault injection are adopted as fault-tolerant approaches. For the weight restriction, a range to which values of weights should belong is determined during the learning, and any weight being outside this range is forced to be either its upper limit or lower limit. A status of a fault occurring is then evoked by the fault injection, and calculating weights is made under this status. The learning based on both of the above approaches surpasses the learning based on either of them in the fault tolerance and/or in the learning time.

Universal Construction on Self-Timed Cellular Automata

Abstract: Computation- and construction-universality in cellular automata (CA), first studied by von Neumann, has attracted steady research efforts, over the years, most employing synchronous CA. Asynchronous cellular automata (ACA), though of interest as most interactions in nature are asynchronous, have not been used for this task, other than by the indirect way of simulating a synchronous CA. In this paper, we propose a universal constructor on a self-timed cellular automaton (STCA), a particular type of ACA, in which cells are divided in four partitions, each with four states.

On improvement in fault tolerance of Hopfield neural networks

Abstract: Hopfield neural networks tolerating weight faults are presented. The network training is made on condition some faults occur. Statuses of such faults are evoked by intentionally injecting faults into the network. The learning using the single-fault injection is shown first. Learning schemes, which are based on the double-fault injection for a couple of weights within a neuron, are then proposed to improve the fault tolerance further. Experimental results show that the learning using the random-double-fault injection allows us to complete the reasonably dependable network with the acceptable length of the learning time. In addition, the proposed schemes make the network robust against the input noise.

An application of self organizing map to detection of confused blood samples

Abstract: We propose a cost-aware method of detecting blood samples confused among different patients, using self organizing maps. The map consists of a cluster reacting to confused data and that reacting to non-confused data. It is completed so that the latter cluster can become larger than the former, and allows reducing drastically the number of samples wrongly judged to be retested.

Reinforcement learning using chaotic exploration in maze world

Abstract: In reinforcement learning, trial and error called an exploration plays an important role. As a generator for exploration, it seems to be familiar to use the uniform pseudorandom number generator. However, it is known that chaotic source also provides a random-like sequence as like as stochastic source. In this research, we propose an application of the random-like feature of deterministic chaos for a generator of the exploration. As a result, we find that the deterministic chaotic generator for the exploration based on the logistic map gives better performances than the stochastic random exploration generator in a nonstationary shortcut maze problem. In order to understand why the exploration generator based on the logistic map shows the better result, we investigate the learning structures obtained from the two exploration generators.

On Retrieval of Lost Functions for Feedforward Neural Networks Using Re-Learning

Abstract: This paper proposes the re-learning for feedforward neural networks where weight faults would occur. The sequences of target outputs are encoded by means of single-parity-check codes so that a single-bit error caused by the faults can be on-line detected at the output layer. The re-learning is made every time a network produces the error, and its lost function is retrieved. The proposed scheme can easily achieve high MTTF (Mean Time To Failure).