Restricted Boltzmann machines (RBMs) are bipartite structured statistical neural networks and consist of two layers. One of them is a layer of visible units and the other one is a layer of hidden units. In each layer, any units do not connect to each other. RBMs have high flexibility and rich structure and have been expected to applied to various applications, for example, image and pattern recognitions, face detections and so on. However, most of computational models in RBMs are intractable and often belong to the class of NP-hard problem. In this paper, in order to construct a practical learning algorithm for them, we employ the Kullback-Leibler Importance Estimation Procedure (KLIEP) to RBMs, and give a new scheme of practical approximate learning algorithm for RBMs based on the KLIEP.

View full abstract

In this paper, quasi-ARX wavelet network (Q-ARX-WN) is proposed for nonlinear system identification. There are mainly two contributions are clarified. Firstly, compared with conventional wavelet networks (WNs), it is equipped with a linear structure, where WN is incorporated to interpret parameters of the linear ARX structure, thus Q-ARX-WN prediction model could be constructed and it is easy-to-use in nonlinear control. Secondly, guidelines for network construction are well considered due to the introduction of WNs, and Q-ARX-WN could be represented in a linear-in-parameter way. Therefore, linear support vector regression (SVR) based identification scheme may be introduced for the robust performance. Moreover, in adaptive control procedure, only linear parameters are needed to be adjusted when sudden changes have happened on the nonlinear system, thus the controller can track reference signal quickly. The effectiveness and robustness of the proposed nonlinear system identification method are validated by applying it to identify a real data system and a mathematical example, and an example of nonlinear system control is given to show usefulness of the proposed model.

View full abstract

The Inverse function Delayed model (ID model) is a neuron model with negative resistance dynamics. The negative resistance can destabilize local minimum states, which are undesirable network responses. The ID network can remove these states. Actually, we have demonstrated that the ID network can perfectly remove all local minima with N-Queen problems or 4-Color problems, where stationary stable states always give correct answers. However this method cannot apply to Traveling Salesman Problems (TSPs) or Quadratic Assignment Problems (QAPs). Meanwhile, it is proposed that the TSPs are able to be represented in terms of the quartic form energy function. In this representation, the global minimum states that represent correct answers and the local minimum states are separable clearly, thus if it is applied to the ID network, it ensures that only the local minimum states are destabilized by the negative resistance. In this paper, we aim to introduce higher order connections to the ID network to apply the quartic form energy function. We apply the ID network with higher order connections to the TSPs or QAPs, and show that the higher order connection ID network can destabilize only the local minimum states by the negative resistance effect, so that it obtains only correct answers found at stationary stable states. Moreover, we obtain minimum parameter region analytically to destabilize every local minimum state.

View full abstract

Quantum computation algorithms indicate possibility that non-deterministic polynomial time problems can be solved much faster than classical methods. We have proposed a neuromorphic quantum computation algorithm based on adiabatic quantum computation, in which an analogy to an artificial neural network is considered in order to design a Hamiltonian. However, in the neuromorphic AQC, the relation between its computation time and success rate has not been clear. In this paper, we study residual energy and the probability of correct answers as a function of computation time. The residual energy behaves as expected from the adiabatic theorem. On the other hand, the success rate strongly depends on energy level crossings of excited states during Hamiltonian evolution. The results indicate that computation time must be adjusted according to a target problem.

View full abstract

In this paper, we implement a model of an electric fish, Eigenmannia, that detects frequency differences between the individuals, on analog CMOS circuits. The circuit's fundamental function is equivalent to a conventional “phase frequency comparator”. The circuit consists of five elemental cell units that implement neural networks of the electric fish. Using a simulation program of integrated circuit emphasis (SPICE), we demonstrate that the proposed circuit can detect the frequency difference.

View full abstract

Hospedales et al. have recently proposed a neural network model of the “vestibulo-ocular reflex” (VOR) in which a common input was given to multiple nonidentical spiking neurons that were exposed to uncorrelated temporal noise, and the output was represented by the sum of these neurons. Although the function of the VOR network is equivalent to pulse density modulation, the neurons' non-uniformity and temporal noises given to the neurons were shown to improve the output spike's fidelity to the analog input. In this paper, we propose a CMOS analog circuit for implementing the VOR network that exploits the non-uniformity of real MOS devices. Through extensive laboratory experiments using discrete MOS devices, we show that the output's fidelity to the input pulses is clearly improved by using multiple neuron circuits, in which the non-uniformity is naturally embedded into the devices.

View full abstract

In this paper we study the generation of an ill-conditioned integer matrix A=[a_ij] with |a_ij|≤µ for some given constant µ. Let n be the order of A. We first give some upper bounds of the condition number of A in terms of n and µ. We next propose new methods to generate extremely ill-conditioned integer matrices. These methods are superior to the well-known method by Rump in some respects, namely, the former has a simple algorithm to generate a larger variety of ill-conditioned matrices. In particular we propose a method to generate ill-conditioned matrices with a choice of desirable singular value distributions as benchmark matrices.

View full abstract

Two random number generation methods based on regular and chaotic sampling of chaotic waveforms are introduced. IC truly random number generators based on these methods are also presented. Simulation and experimental results, verifying the feasibilities and correct operations of the circuits, are given. Numerical models for the proposed TRNG designs have been developed leading the realization of the random number generator circuits. Moreover, a feedback strategy including offset and frequency compensation circuits have been developed in order to maximize the statistical quality of the output sequence and to be robust against external interference, parameter variations and attacks. Prototype chips have been fabricated by using HHNEC's 0.25 µm eFlash process with a supply voltage of 2.5V, which feature throughput in the order of a few Mbps and fulfill the NIST-800-22 statistical test suites for randomness without post-processing.

View full abstract