In this study, a novel ergodic sequential logic (SL) central pattern generator (CPG) for the realization of multiple gaits of a six-legged robot is presented. First, a novel generalized ergodic SL oscillator and its theoretical analysis method are presented. Using the theoretical analysis method, it is shown that the ergodic SL oscillator can realize oscillations with various amplitudes by adjusting its parameters. Second, a novel ergodic SL CPG consisting of a network of the ergodic SL oscillators is presented. The SL CPG can be classified into four classes (i.e., ergodic CPG of ergodic SL oscillators, ergodic CPG of regular SL oscillators, regular CPG of ergodic SL oscillators, and regular CPG of regular SL oscillators) depending on properties of their driving signals. Detailed analyses show that the ergodic CPG of ergodic SL oscillators is most suited to be used as the CPG for controlling the six-legged robot. Third, the presented CPG is implemented by a filed programmable gate array. Experiments validate that the presented CPG can realize six-legged and four-legged gaits of the six-legged robot. Furthermore, it is shown that the presented CPG is much more hardware-efficient compared to some commonly used conventional CPGs.

In recent years, soft errors on VLSI have increased due to miniaturization, higher integration, and lower supply voltages. A soft error is a temporary error caused by the impact of high-energy particles, such as neutrons from cosmic rays, on VLSI. Recent studies have proposed flip-flops (FFs) tolerant to soft errors occurring on VLSI system scaling continues. Non-volatile (NV) FFs have been studied to reduce the power consumption of battery-powered devices. This paper has presented a design of NV-FFs with soft-error tolerant capability using the DICE (dual interlocked storage cell) structure, which is a soft-error tolerant structure.

In this study, a novel ergodic sequential logic neuron model is presented for the purpose of application to neural prosthesis. It is shown that the presented neuron model can realize various neuron-like responses such as spiking and bursting. Then, a biologically plausible conductance-based Hodgkin-Huxley-type auditory neural network model of a female cricket is presented as a target for virtual trial of the neural prosthesis. It is shown through the virtual neural prosthesis trial that the presented ergodic neuron model can recover the responses of the biologically plausible neural network that lost proper responses due to damage to a neuron model. Furthermore, it is shown that the presented neuron model is more hardware-efficient compared to a commonly used digital processor neuron model. Finally, the significance and potential impact of this study are discussed.

In an earthquake or other disaster, we need to quickly share information on the damage and evacuation status of the affected areas. However, the information management systems used currently are centralized and may not be able to share information on the disaster area quickly due to traffic congestion and server problems. To solve this problem, a decentralized management system using blockchain has been proposed. However, blockchain has the problem that the processing time depends on the number of transactions. In this paper, we assume a prioritized management system for disaster information, and evaluate the delay time of the blockchain using queueing theory. In addition, we analyze the mean sojourn time for other information to compare the mean sojourn time for disaster information.

Out-of-distribution (OOD) data detection is a key challenge in securing artificial intelligence (AI) applications. To address this challenge, this study focused on Bayesian neural networks (BNNs), which can estimate uncertainty in AI. This study performed two major verifications to validate the effectiveness of BNNs in OOD detection One was the verification of the basic OOD detection performance of BNN, and the other was that of the effectiveness of the image data preprocessing method proposed herein to improve the performance. The results showed that all BNNs trained on five benchmark data sets exhibited high OOD data detection performance. Further, the conditions under which BNNs can achieve higher performance were identified. Subsequently, the proposed method was shown to increase the OOD data detection performance on certain data.

The challenge with existing JPEG artifact removal methods is that the restored images are excessively smoothed, resulting in low perceptual quality. This is due to the use of the L1 or L2 norm as the loss function, which has a low correlation with perceptual quality. A common approach to solving this problem, the adversarial generative network (GAN), makes the learning process unstable. In this paper, we propose a new perceptual JPEG artifact removal method by using the weighted sum of multiple IQAs as a loss function. Furthermore, we modify the upscaling architecture of the existing method to prevent periodic artifacts caused by changes in the loss function. Experimental results show that the proposed method significantly improves the perceptual quality of artifact removed images quantitatively and qualitatively compared to the existing baseline.

Image steganography is a technique for embedding secret messages in images. SteganoGAN, one of the previous methods, uses generative adversarial networks and achieves high payload when the input is a noise-free image. However, real images contain real image noise (RIN) generated during the image acquisition process, which can degrade the performance of SteganoGAN. We propose a RIN-aware image steganography that uses a real image denoising method as preprocessing and modifies the loss function of SteganoGAN. These modifications encourage the proposed method to embed a secret message that simulates RIN into a pseudo-noise-free image obtained by denoising. Experimental results show that the proposed method improves the trade-off between image quality and payload compared to the previous method. We validate the statistical and neural steganalysis and JPEG robustness, showing that the proposed method has reasonable detection avoidance capability and higher compression tolerance than the conventional methods.

We examined the memory capacity (MC) characteristics of a reservoir neural network (RNN) that uses an extended chaotic neural network (ExCNN) model as the reservoir layer. Furthermore, we propose a novel prediction performance measure based on the long tail property of MC. To design RNN using the ExCNN model, we derived the echo state property of the ExCNN model. To evaluate the effect of exponential local memory terms of the ExCNN model on prediction performance, we performed a closed-loop one-step prediction of a periodic random sequence. The results indicate that RNN using the ExCNN model with high prediction performance exhibits a distinctive long tail in MC characteristic curve with respect to a calculation delay.

Delay-based reservoir computing is a physical implementation scheme of machine learning and could lead to more energy-efficient and faster processing than traditional digital computing methods. Reservoir computing based on echo-state networks has a well-known trade-off between memory and nonlinearity. However, this trade-off has not been extensively studied in the context of delay-based reservoir computing, and methods for adjusting memory and nonlinearity have not yet been explored. In this study, we show that the delay-based reservoir computing exhibits a trade-off relationship, where a reservoir with a small delay time produces higher nonlinearity and lower memory, and one with a large delay time produces opposite properties. Our results indicate that nonlinearity and memory can be controlled by varying the feedback delay time in delay-based reservoir computing. Moreover, we show that a delay-based reservoir with a small delay time can produce nodes with different nonlinearity and memory properties. These findings have a potential to significantly enhance the performance and efficiency of delay-based reservoir computing. We investigate the reservoir computing performance of two chaotic time-series prediction tasks. A small delay reservoir produces a high prediction performance for multivariate time-series analysis.The proposed method allows for effective control of reservoir nonlinearity without the need to tune physical nodes.

Hippocampus integrates events and places information critical for episodic memory. Memory-based hippocampus-inspired model (MBHIM) is a hippocampus-inspired model suitable for VLSI implementation. This paper proposes an efficient spatial representation scheme for MBHIM with VLSI implementation. Each layer represents a space of a different size to represent and store spatial information at different resolutions. This scheme can achieve the representation of large spaces with high spatial resolutions, even when the number of memory cells per layer is limited. We conducted an experiment using a VLSI system of MBHIM to demonstrate the effectiveness of the proposed scheme.

A spatiotemporal contextual learning network (STCLN) model is suitable for edge applications. To implement the STCLN model as a small analog asynchronous integrated circuit with low power consumption, this study proposes to introduce the excitatory and inhibitory synapses into the STCLN model which can be implemented by an event-driven spiking neural network (SNN) circuit. Through the performance evaluation of the proposed STCLN model for SNN, the optimal ratio of the excitatory and inhibitory synapse is derived.

The field of neuromorphic computing has experienced remarkable growth driven by the need to overcome the limitations of traditional von Neumann architecture in handling big data, IoT, and AI tasks. This paper provides an overview of recent advancements in brain-inspired computing, particularly focusing on the integration of various memristive materials, including metal oxides, perovskites, and 2D materials, into neuromorphic hardware. The evolution of artificial neural network (ANN) technology, ranging from perceptron to deep neural networks (DNNs) and spiking neural networks (SNNs), is discussed, emphasizing the potential of SNNs for energy-efficient hardware implementation. Challenges in integrating memristors, especially 2D material-based memristors, into SNNs are highlighted, along with recent developments in neuromorphic hardware utilizing memristors and complementary metal-oxide-semiconductor (CMOS) technology. Through simulations and experimental demonstrations, researchers have shown the feasibility of using memristors for implementing artificial neurons and synapses, paving the way for efficient and scalable neuromorphic computing systems.

A single-celled amoeba can solve the traveling salesman problem through its shape-changing dynamics. In this paper, we examine roles of several elements in a previously proposed computational model of the solution-search process of amoeba and three modifications towards enhancing the solution-search performance. We find that appropriate modifications can indeed significantly improve the quality of solutions. It is also found that a condition associated with the volume conservation can also be modified in contrast to the naive belief that it is indispensable for the solution-search ability of amoeba. A proposed modified model shows much better performance.

Image anomaly detection is a crucial task in computer vision, where convolutional neural networks (CNN) often deliver exceptional performances. Hardware implementation of machine learning models is also important for achieving inference speed-up and power savings. However, the massive number of CNN parameters poses challenges for hardware implementation. This study introduces reservoir computing (RC) to create a compact image processor without training, thereby enabling scalable deployment. Our proposed bidirectional 2-dimensional reservoir computing (BiRC2D) is a feature extractor based on RC. Experiments conducted on the MVTec AD dataset, a benchmark dataset for real-world anomaly detection task, validated the efficacy of BiRC2D when integrated into the patch distribution modeling (PaDiM) framework. The mean intersection over union (mIoU) score from PaDiM with BiRC2D outperformed or was comparable to the mIoU score from PaDiM with ResNet-50 in several categories while reducing the parameter count by up to 98%.

In this study, we propose a technique for replicating physical reservoirs using a small-scale digital calibration reservoir. Recently, physical reservoirs, which comprise the physical systems for reservoir computing and computational methods for time-series pattern recognition, have attracted considerable attention. However, replicating physical reservoirs is challenging because of the individual variations that occur when physical systems are used. We introduce a small-scale digital reservoir that absorbs individual variations, compensates for them, and successfully replicates the performance of a physical reservoir.

Three-dimensional physical reservoir is made by filling Ag₂S particles in a cylindrical space with input/output electrodes on both its top and bottom surfaces. Functions of nonlinear transformation and short-term memorization of input signals are achieved by the change in biased distribution of Ag⁺ cations in each Ag₂S particle, enabling the simple fabrication of the three-dimensional network just by filling Ag₂S particles into the cylindrical space. Benchmarking with the short-term memory task and with the parity-check task revealed that the three-dimensional reservoir of Ag₂S particles shows better performance than a two-dimensional reservoir of Ag₂S particles, even with use of the same number of electrodes.

Neuromorphic devices have diverse potential applications, such as brain-inspired computers and promising high-performance arithmetic systems with power saving. Reservoir computing (RC), a type of recurrent neural network (RNN), achieves learning by adjusting the weights between the intermediate and output layers. Time-delay reservoir computing introduces a delay and creates virtual nodes within the middle layer. The Ag/Ag₂S nanoparticles function as nonlinear electrical devices, following the atomic switch principles of the time-delay system. Voice recognition was performed with 87.81% accuracy when six different people pronounced the same number, and 80.18% when the same person pronounced ten different numbers.

This paper presents a system for processing tactile information using a carbon nanotube (CNT)-polydimethylsiloxane (PDMS) nanocomposite sensor designed for robotic applications. This study introduces an approach for recognizing objects and detecting optimal grasping points using tactile data from a sensor-equipped robotic hand. The sensor is expected to be more efficient than a computerized implementation using sensor dynamics directly in the computation. The experiments demonstrated the system's ability to classify nine types of objects with an accuracy of 83.56% and to discern two grasping points on four object types, achieving a 71.7% success rate.

This study discusses the use of reservoir computing in modeling dynamic systems of machinery for detecting anomalies. Precise modeling of a dynamic system of a machine is essential for detecting anomalies based on the comparison of predicted and actual outputs. To achieve this, motors were used as the subject of study, and three physical systems were constructed by progressively applying loads to their complex dynamic systems. By exploring the parameters of the reservoir and using preprocessed actual data, we successfully captured these physical systems through reservoir computing. Additionally, this study evaluated the predictive accuracy based on specific input patterns and assessed the model response to intentionally created abnormal conditions (manual stopping of the motor). In scenarios involving specific inputs that frequently occur in actual operational environments, the model showed significant discrepancies between predicted and actual values. These results indicate that reservoir computing can detect unexpected dynamic changes and effectively distinguish between normal and abnormal operating conditions. This study confirms that reservoir computing is an effective tool for the accurate modeling of machinery's dynamic systems and for detecting unexpected anomalies in real operational environments.

The Hopfield networks have been successfully applied to associative memories for binary data. The complex-valued Hopfield networks extend the classic Hopfield networks for multilevel data. Several researchers have been attempting further extensions by Clifford algebra. However, the dimensions of Clifford algebras are limited to power of two. More flexible algebraic families are necessary for extensions of high-dimensional Hopfield networks. In the present paper, we propose high-dimensional Hopfield networks using group rings, referred to as group ring-valued Hopfield networks (GRVHNs). The GRVHN is a broad family which provides high-dimensional Hopfield networks of any dimensions. Moreover, it is also shown that many conventional high-dimensional Hopfield networks are included in the GRVHNs.

This paper discusses discretization methods for implementing nonlinear model predictive controllers using Iterative Linear Quadratic Regulator (ILQR). Finite-difference approximations are mostly used to derive a discrete-time state equation from the original continuous-time model. However, the timestep of the discretization is sometimes restricted to be small to suppress the approximation error. In this paper, we propose to use the variational equation for deriving linearizations of the discretized system required in ILQR algorithms, which allows accurate computation regardless of the timestep. The use of the variational equation with a longer timestep can improve control performance, in terms of the optimality of the trajectory and the robustness to measurement noises, by increasing the number of ILQR iterations possible at each timestep in the real-time computation. Case studies of the swing-up control of an inverted pendulum on a cart and a rotary inverted pendulum are provided to demonstrate the effectiveness of the proposed method.

Machine learning methods allow the prediction of nonlinear dynamical systems from data alone. The Koopman operator is one of them, which enables us to employ linear analysis for nonlinear dynamical systems. The linear characteristics of the Koopman operator are hopeful to understand the nonlinear dynamics and perform rapid predictions. The extended dynamic mode decomposition (EDMD) is one of the methods to approximate the Koopman operator as a finite-dimensional matrix. In this work, we propose a method to compress the Koopman matrix using hierarchical clustering. We performed numerical experiments on cart-pole and rope models and compared the results with those of the conventional singular value decomposition (SVD); the results indicate that the hierarchical clustering performs better than the naive SVD compressions.

In order to cope with the ever-increasing demand and volume of data processing, the intelligence has been moved to the edge of the IoT systems. We propose a brain-morphic wireless sensor network (B-WSN), which incorporates reservoir computing in a WSN and performs neural computation. By exchanging pulse signals among neuromorphic nodes, the network exhibits input-dependent dynamics from which information about the inputs is derived. We discuss design options and apply our proposal to estimating the temperature distribution in an area with the B-WSN to verify its performance. Results showed that it could estimate the temperature with the average RMSE of 5.38 and 11.25 degree Celsius in training and testing, respectively.