礒川 悌次郎

Deep echo state network (Deep-ESN) model consists of multiple reservoir layers, that can respond to layer-specific different time-scales. This dynamical characteristic leads to enhance performance of ESN. However, neither the design guidelines for the hyperparameters of the network and individual neurons nor the mechanism to produce the diverse dynamical response have been clarified. In this study, we proposed an approach to generate the dynamical responses with different time-scales in each layer by adjusting the leaking rate of neurons. Through the evaluations time-series prediction task for different leaking rates, multiscale entropy analysis for each reservoir layer, and cross-correlation between adjacent layers, we found that when the leaking rate is set to low, the layer-specific dynamics with different time-scales are generated, as well as a mechanism whereby the signal propagates to subsequent layers with a delay, i.e., queue characteristic. These characteristics produce a high memory capacity. Consequently, diverse responses with delay leads to an enhancement of Deep-ESN functionality.

In this study, we investigate model structures for neural ODEs to improve the data efficiency in learning the dynamics of control systems. We introduce two model structures and compare them with a typical baseline structure. The first structure considers the relationship between the coordinates and velocities of the control system, while the second structure adds linearity with respect to the control term to the first structure. Both of these structures can be easily implemented without requiring additional computation. In numerical experiments, we evaluate these structure on simulated simple pendulum and CartPole systems and show that incorporating these characteristics into the model structure leads to accurate learning with a smaller amount of training data compared to the baseline structure.

<p>In recent ethological studies, the behaviors and interactions of animals have been recorded by digital video cameras and webcams, which provide high functionality at reasonable cost. However, extracting the behavioral data from these videos is a laborious and time-consuming manual task. We recently proposed a novel method for tracking unmarked multiple honeybees in a flat arena, and developed a prototype software named "K-Track". The K-Track algorithm successfully resolved nearly 90% of cases involving overlapped or interacted insects, but failed when such events happened near an edge of a circular arena, which is commonly employed in experiments. In the present study, we improved our K-Track algorithm by comparing the interaction trajectories obtained from forward and backward playing of video episodes. If the tracking results differed between the forward and backward episodes, the trajectory with lower maximum moving distance per frame is chosen. Based on this concept, we developed a new software, "K-Track-kai", and compared the performances of K-Track and K-Track-kai in honeybee tracking experiments. In the cases of 6 and 16 honeybees, K-Track-kai improved the tracking accuracy from 91.7% to 96.4% and from 94.4% to 96.7%, respectively.</p>

© 2018 IEEE. In this paper, an ultrasound-measurement-based method of detecting stenosises in fallopian tubal models is presented. The proposed method employs silicone rubber tubes and custom phantoms made of polymer gels as models of fallopian tubes. When the targets are the former (or latter), stenosis models are formed by lumps of polyester sewing threads (or polymer gels with high reflection intensity). The proposed method copes with detecting stenosis models as classification problems by applying support vector machines (SVMs for short). It prepare data for SVM learning and stenosis detection from waveforms in the range specified by two time points at which waves are probably reflected by inner upper and lower surfaces of tubal models. The data preparation is conducted by generating the frequency distribution of the number of regular short intervals into which this range is divided versus normalized values associated with amplitude of the waveforms in such intervals. When it is explored whether given data are prepared from a tube with the stenosis model, a discrimination model constructed by SVM learning works as a two-class classifier. Using the model as a multi-class classifier makes it possible to distinguish data of tubes with stenosis models from each other. It is revealed from experimental results that the proposed method has substantially high capability in detecting stenosis models.

© 2017 IEEE. In recent studies, researchers can easily record the behaviors of animals by digital video cameras, which provide high functionality at reasonable cost. However, it is a laborious and time-consuming manual task for them to extract the useful behavioral data from these videos. We recently proposed a tracking method for unmarked multiple honeybees in a flat arena, named the 'K-Track' algorithm. The algorithm can successfully identify and track nearly 90% of interaction cases of targets. In this study, we proposed an improved method for the existing K-Track algorithm by tracking results using backward-play movie. If the tracking results differed between the forward and backward episodes, one of them had probably resulted from correct tracking. Therefore, by comparing the forward and backward trajectories of the same interaction, we assumed that there is the potential for an increase in tracking accuracy. In the experiments, K-Track using backward movies successfully tracked four out of five situations that was failed by original K-Track and we confirmed that the method has the potential of improved tracking accuracy.

© 2017 IEEE. In this paper, a method of calculating inner and outer diameters of fallopian tubes is presented, provided that silicone rubber tubes are models for fallopian tubes. It is based on the ultrasound measurement using a single probe of which the nominal frequency is 5 MHz, and assumes that ultrasound waves emanating from the probe are reflected at the top outer surface, surfaces between water and silicone rubber, and bottom outer surface of the tube. It obtains envelope curves associated with the reflected waves, by applying Hilbert transformation to the waves. It next estimates top four maximal values on the envelope curves to acquire time points when the wave reflections occur. The inner and outer diameters of the tubes are easily estimated by substituting the above time points in simple formulas. Experimental results reveal that the proposed method can achieve substantial accuracy in estimating inner and outer diameters of the target tubes.

In this paper, a method of estimating binding time is proposed for ophthalmological outpatients. Binding time equals the sum of transit time and waiting time at a hospital. It determines the number of outpatients virtually arriving and their arrival time intervals according to the gamma distribution and the exponential distribution. It then assigns examinations to outpatients virtually arriving, referring to their appointment types probabilistically determined. Ten examination rules are established to simulate the passage of time required for examinations and consultation. The proposed method accesses database associated with public transport via some Application Programming Interface, and finally provides transit time and waiting time estimated by the simulation. Experimental results on estimating waiting time show that the average daily difference between actuality and estimation is acceptable.

This paper presents a 3-state asynchronous cellular automaton (CA) that requires merely two transition rules to achieve computational universality. This universality is achieved by implementing Priese’s delay-insensitive circuit elements, called the E-element and the K-element, on the cell space of a so-called Brownian CA, which is an asynchronous CA containing local configurations that conduct a random walk in the circuit topology.

Hebbian learning rule is well known as a memory storing scheme for associative memory models. This scheme is simple and fast, however, its performance gets decreased when memory patterns are not orthogonal each other. Pseudo-orthogonalization is a decorrelating method for memory patterns which uses XNOR masking between the memory patterns and randomly generated patterns. By a combination of this method and Hebbian learning rule, storage capacity of associative memory concerning non-orthogonal patterns is improved without high computational cost. The memory patterns can also be retrieved based on a simulated annealing method by using an external stimulus pattern. By utilizing complex numbers and quaternions, we can extend the pseudo-orthogonalization for complex-valued and quaternionic Hopfield neural networks. In this paper, the extended pseudo-orthogonalization methods for associative memories based on complex numbers and quaternions are examined from the viewpoint of correlations in memory patterns. We show that the method has stable recall performance on highly correlated memory patterns compared to the conventional real-valued method.

A quaternionic extension of feed forward neural network, for processing multi-dimensional signals, is proposed in this paper. This neural network is based on the three layered network with random weights, called Extreme Learning Machines (ELMs), in which iterative least-mean-square algorithms are not required for training networks. All parameters and variables in the proposed network are encoded by quaternions and operations among them follow the quaternion algebra. Neurons in the proposed network are expected to operate multidimensional signals as single entities, rather than real-valued neurons deal with each element of signals independently. The performances for the proposed network are evaluated through two types of experiments: classifications and reconstructions for color images in the CIFAR-10 dataset. The experimental results show that the proposed networks are superior in terms of classification accuracies for input images than the conventional (real-valued) networks with similar degrees of freedom. The detailed investigations for operations in the proposed networks are conducted.

Asynchronous Boolean totalistic cellular automata have recently attracted attention as promising models for the implementation of reaction-diffusion systems. It is unknown, however, to what extent they are able to conduct computation. In this paper, we introduce the so-called non-camouflage property, which means that a cell’s update is insensitive to neighboring states that equal its own state. This property is stronger than the Boolean totalistic property, which signifies the existence of states in a cell’s neighborhood, but is not concerned with how many cells are in those states. We argue that the non-camouflage property is extremely useful for the implementation of reaction-diffusion systems, and we construct an asynchronous cellular automaton with this property that is Turing-complete. This indicates the feasibility of computation by reaction-diffusion systems.

The aim of this paper is to investigate storing and recalling performances of embedded patterns on associative memory. The associative memory is composed of quaternionic multistate Hopfield neural network. The state of a neuron in the network is described by three kinds of discretized phase with fixed amplitude. These phases are set to discrete values with arbitrary divide size. Hebbian rule and projection rule are used for storing patterns to the network. Recalling performance is evaluated through storing random patterns with changing the divide size of the phases in a neuron. Color images are also embedded and their noise tolerance is explored.

A qubit neural network (QNN) is a neural network that incorporates the quantum computing and representation. QNN is constructed from a set of qubit neuron model, of which internal state is a coherent superposition of qubit states. This paper evaluates the performance of QNN through a prediction of well-known Lorentz attractor, which produces chaotic time series by three dynamical systems. The experimental results show that QNN can predict time series more precisely, compared with conventional (real-valued) neural networks.

Grouping behavior, such as bird flocking, terrestrial animal herding, and fish schooling, is one of well-known emergent phenomena. Several models have been proposed for describing grouping behaviors, and two types of models can be defined: rule-based model and learning-based model. In rule-based models, each agent in a group has fixed interaction rules with respect to other agents. On the other hand, agents in learning-based model acquire their rules by the interactions of other agents with a learning scheme such as Q-learning. In this paper, we adopt quantities obtained from trails of agents, in order to investigate the properties for grouping behaviors of agents. We also evaluate rule-based and learning-based models by using these quantities under the environments with and without predatory agents.