In this study, an ergodic sequential logic central pattern generator model was developed to demonstrate the capability of producing various gait patterns. A simple parameter search method enabled the model to acquire typical hexapod gaits such as tripod, tetrapod, and wave gaits. The model was then implemented in a field-programmable gate array (FPGA), and experiments validated that a hexapod robot equipped with the FPGA walked appropriately. It was also shown that the model could be implemented with fewer circuit elements and consumed less power for operation than the conventional CPG model. These results suggest that the model is suitable for gait controllers of neuromorphic robots.

A complex-valued Hopfield associative memory (CVHAM) has been studied as a multistate model of neural associative memory. The weak noise tolerance is a disadvantage of CVHAM. A symmetric complex-valued Hopfield associative memory (SCVHAM), a modification of CVHAM, provides much better noise tolerance than CVHAM, though the projection rule is not applicable. In this work, a hybrid complex-valued Hopfield associative memory (HCVHAM) is proposed. An HCVHAM is a combination of a CVHAM and an SCVHAM, and takes both advantages of CVHAMs and SCVHAMs. The quaternion projection rule, which was first introduced to rotor Hopfield associative memories, is applied to an HCVHAM. Computer simulations show that the HCVHAMs provide excellent noise tolerance.

In recent years, artificial intelligence (AI) has attracted attention to edge AI, which operates in an offline environment without the cloud and emphasizes response time. The demands of training neural networks at the edge have been actively discussed, but the problem is that the large amount of memory and computation required exceeds the limits of resources available at the edge. Much memory is allocated to the optimizer to hold and update model parameters, and advanced optimizers require memory to store additional parameters such as moments (past gradient information) for each parameter. Therefore, this research aims to reduce the amount of memory allocated to the optimizer and realize edge AI by using Holmes, an optimizer designed for implementation at the edge. In this study, we verify the applicability of Holmes to recurrent neural networks (RNNs), a variant of neural networks commonly used for time-series data. In a previous study, Holmes was proven to achieve sufficient accuracy using the MNIST dataset, but its use in RNNs has not been well confirmed. The difficulty in applying Holmes to RNNs is that the layers are effectively deeper due to the characteristic of RNNs having feedback paths. As we proceeded step by step from relatively easy verification to evaluation, we discovered the possibility that Holmes could be applied to RNNs. We present several validations we have performed, mainly on training of function prediction and language processing, which we compare and evaluate with other optimizers in terms of training accuracy. Through the proof-of-concept implementation and evaluation of Holmes in RNNs, we expand the possibilities toward edge training of RNNs. The study shows Holmes' potential for efficient edge AI applications, enabling resource-constrained devices to handle complex RNN tasks with accuracy comparable to traditional optimizers.

Developing hardware for Artificial intelligence (AI) training is vital. A hardware-oriented optimizer, named Holmes enables faster training with a smaller memory footprint. This study developed a hardware architecture that incorporates Holmes and benefits from parallelization and pipelining to achieve significant throughput improvement. We determined the required bit width for training and used it the architecture evaluation. We investigated scalability and the effectiveness of both Holmes and pipelining. The results proved the linear scalability of the memory footprint over the model size, reduction of the memory footprint by utilizing Holmes, drastic increase in throughput by pipelining and faster computing.

Hippocampal neurons that representing the animal's self-location are called “place cells”. During the maze task, sequential firing of place cells during exploration is reproduced as fast synchronous firing during rest in rodents, which is called hippocampal replay. Reverse replays firing in the reverse order of the path occur, and replays of pathways not experienced by the rats also occur. The majority of the replays of pathways not experienced represent shortcuts to reward points in the maze. These observations suggest that hippocampal replay contributes to spatial learning. However, there is a lack of spiking neural networks with mechanisms for the emergence of replays of shortcut pathways. In this study, we constructed a spiking neural network of the hippocampal CA3 region. We propose a possible mechanism by which shortcut replays are acquired through synaptic plasticity during a figure-eight maze exploration. In the proposed hippocampal spiking network, the combination of forward and reverse replays results in shortcut replays.

Edge computing methods, particularly federated learning (FL), have gained significant attention because of their ability to share contributions to model training without exposing the training data to other users. Although conventional FL focuses on a single-task data distribution, it is unclear whether similar learning can be achieved for different tasks. We explore a learning architecture that supports multitask learning through partial model sharing. To demonstrate its effectiveness, we propose the common bases hypothesis, suggesting that the efficient sharing of common representations among tasks is possible. Using singular value decomposition to fix a subspace of the trained weights during learning, our results indicate that FL with partial model sharing for multiple tasks is feasible because of its ability to learn similar representations.

In our previous study, we proposed Optimizer based on Spiking Neural-oscillator Networks (OSNNs) which are one of the deterministic particle swarm optimization methods. OSNNs are based on dynamics of a spiking neural-oscillator (spiking oscillator) network which consists of plural integrate-and-fire type neurons. In OSNNs, each particle consists of plural spiking oscillators which are coupled with other spiking oscillators by a network topology. The Ring 1-way network topology, which is one of the simple ring-type networks, is introduced into the network topology for OSNNs. The Ring 1-way network can affect search performances of OSNNs, and OSNNs with the Ring 1-way network (Ring OSNNs) have better search performances than Optimizer with Uncoupled Spiking Neurons (OUSNs) in which each spiking oscillator does not connect to the others. In particular, search performances of Ring OSNNs are clearly superior to those of OUSNs in solving separable and multimodal problems. However, the reasons why the Ring 1-way network can improve search performances of OSNNs have not been clarified. In this study, we theoretically analyzed the search dynamics of the Ring 1-way network. In addition, in order to demonstrate analytical results, we conduct numerical simulations by comparing Ring OSNNs with OUSNs. We further discuss the simulation results and clarify effectiveness of the Ring 1-way network.

In edge computing, data are processed on the device side in advance, often by means of reservoir computing. Ensemble Kalman filters can be used to improve the learning processes of reservoir computing methods. In this study, we designed and validated an architecture for this approach, where we implemented techniques such as parallel computation by initiating streaming processes, reducing dividers, and accumulating random numbers. The validation results demonstrate that the proposed architecture reduces the time and resource costs of computation while maintaining a sufficient estimation accuracy. These results may facilitate the implementation of AI methods on a small scale.

This study investigates the use of the encoder-decoder model and its application to generative artificial intelligence (AI) for learning on edges. Current generative AI mainly uses a machine learning model called Transformer. However, the core of this model is the existing encoder-decoder model and the attention mechanism. Therefore, by focusing on the encoder-decoder model, we implement and evaluate a sequence transformation model called Sequence to Sequence (Seq2seq) to achieve a generative AI that can be trained on edges. We evaluate the model's performance on an arithmetic task, which is needed to gain a common representation between the input and output. The implementation and evaluation demonstrate the ability to perform the sequence transformation tasks. Throughout the study, we show the prospect of generative AI that can perform on edges.

This study, we discusses the topological type of nonautonomous systems with periodic solutions when the time variable t increases negatively. When linear approximation holds near a fixed point of the Poincaré map, we confirm that the bifurcation points where the fixed point becomes nonhyperbolic are invariant regardless of the time direction however, the stability of the fixed point is changed. Consequently, we show that two-dimensional bifurcation diagrams obtained by the brute-force method give different results for positive and reversal-time variable systems; however, the bifurcation curves are identical. The inverted time variable system is useful for visualizing the completely unstable fixed point, because the repeller can be observed as an attractor. Furthermore, in certain models, chaotic attractors with a wide parameter range exist in reversal time variable systems.

Participatory sensing, which enables extensive sensing with low implementation costs, is expected to be utilized in various fields, such as traffic monitoring, infrastructure management, environmental monitoring, and smart cities. Participatory sensing requires users to pay for battery consumption, communication costs, and other costs when sensing. Therefore, the system needs to provide an incentive mechanism that rewards users within the budget they possess to motivate them to actively participate in sensing. While many existing studies discuss incentive mechanisms that select the optimal users within the budget and pay rewards, there is a lack of discussion regarding user departure. Moreover, our previous work considering user departure assumes that users will transfer sensing costs to the system. However, in the real world, the method of users sending sensing costs to the system is difficult to achieve because of privacy concerns. In this paper, we propose an incentive mechanism that considers the user departure by introducing the non-selection rate of users, which represents the proportion of times a user was not selected by the system, to select the optimal users, to maintain the number of users and data quality.

The mathematical model of a cell population interacting through a common pool (external environment) is a representative model of spatial-independent cell differentiation. This model shows an interesting phenomenon: the ratio of cell differentiation (differentiation ratio) is restricted to a specific range. However, the number of variables increases as the number of cells increases, making the analysis difficult for large cell populations. Therefore, this study takes “symmetry” of the cell population model into account, introducing two different reduced models to decrease the number of variables. Through simulations and bifurcation analyses, several aspects of the differentiation ratio regulation dynamics in the original model are revealed.

Vertebrate skin patterns such as stripes and spots are formed autonomously. In this paper, we found that 2-dimensional patterns not seen in the continuous and the previous Cellular Automata (CA) models appear by imposing nonlinearity on the two-species interaction of CA model that forms the pattern. This extension is a framework that can be used for other CA models and can be a useful tool to investigate the interaction details regarding pattern formation.

This study focuses on accelerationg the learning process in complex-valued Hopfield neural network (CVHNN) models, an extension of the traditional Hopfield model with complex numbers in neuron states and synaptic weights. A major challenge for CVHNN models is the slow convergence of the gradient descent learning rule (GDLR). To solve this problem, we implemented a combination of parallel computation utilizing graphic processing units and complex-domain optimization techniques to boost the speed and stability of GDLR execution. This approach accelerated learning, enhanced memory capacity, and increased tolerance to noise.