This paper studies realization of desired digital spike-trains based on a simple evolutionary algorithm. First, the dynamics of spike-trains is visualized by a digital spike map. The map is defined on a set of points and is represented by a characteristic vector of integers. Second, in order to realize desired spike-trains, we present a simple evolutionary algorithm that aims at optimization of the characteristic vector for a cost function. We also present a simple method to super-stabilize desired spike-trains. Third, in order to implement the digital spike map, we introduce a digital spiking neuron consisting of two shift registers and a wiring between them. An elementary FPGA-based circuit is presented and a super-stable spike-train is demonstrated.
In this work, we study the presence of stochastic resonance (SR) in a hardware pulse-type cell body model. SR is a phenomenon that uses noise to boost weak periodic signals. For this purpose, we design a random noise source and include it in the model because it has been demonstrated that adding random noise to a pulse-type cell body model significantly improves its output signal-to-noise ratio (SNR). The noise also reduces the voltage needed for circuit oscillation. Finally, a decrement in the delay time is observed in the presence of noise.
This paper describes a novel quasi-Newton (QN) based accelerated technique for training of neural networks. Recently, Nesterov's accelerated gradient method has been utilized for the acceleration of the gradient-based training. In this paper the acceleration of the QN training algorithm is realized by the quadratic approximation of the error function incorporating the momentum term as Nesterov's method. It is shown that the proposed algorithm has a similar convergence property with the QN method. Neural network trainings for the function approximation and the microwave circuit modeling problems are presented to demonstrate the proposed algorithm. The method proposed here drastically improves the convergence speed of the conventional QN algorithm.
The gene affects various behaviors of animals such as circadian rhythm, courtship behavior, motor behavior, visual behavior, and learning. The circadian rhythm is a biological rhythm having a period of almost 24 hours, which is sometimes called internal clock or biological clock. In this paper, a novel asynchronous cellular automaton model of a gene network is proposed, where its vector field is designed based on an ordinary differential equation gene network model. It is shown that the proposed model can reproduce typical phenomena (e.g., oscillations, mutual synchronization, locking to light stimulation, and related bifurcations) observed in the differential equation gene network model. It is also shown that the proposed model can be implemented on an FPGA by using much less hardware resource compared to the differential equation model
This paper presents sigma-delta (SD) domain bandpass and band-elimination wave filters modeled after analog distributed parameter filters (ADPF) built up of coupled transmission lines. The bandpass and band-elimination filters have the following advantages: (i) Design techniques of ADPFs can be applied directly and (ii) shift registers and sorting networks both of which are spatially uniform in structure and local in internal connections are the main parts of the SD domain filters. We designed an SD domain bandpass filter and confirmed by computer simulation that it possessed the same frequency response characteristics with the theoretical characteristics of the reference ADPF. We also designed a band-elimination filter from which coefficient multipliers were removed and for which shift registers of different lengths were employed to simplify further the filter structure. Finally, we investigate the fault-tolerant capability of a bandpass block for transient errors of internal circuit cells. With the help of computer simulation, we found that the bandpass block did not fail in operation and did not output high surge noise in the presence of the errors although its output quality decreased.
Wind power has increased rapidly worldwide in recent years. For power system and wind farm operators, it becomes important to understand the smoothing effects of aggregated power, and the temporal and spatial scales at which smoothing is achieved. Here, we propose a new smoothing index for wind power based on the so-called Koopman Mode Decomposition (KMD). KMD decomposes spatio-temporal data on complex wind power into modes oscillating with single frequencies. We show that the proposed smoothing index is regarded as a generalization of a previously proposed index based on power spectral densities. We then look at smoothing of wind power in Japan on a large-scale by incorporating highly-resolved wind prediction data from the Cloud Resolving Storm Simulator (CReSS). In particular, we consider six regions in northern Honshu (the largest island of Japan) as a test case. By applying the proposed index to simulated wind power, we show how the smoothing improves by distributing wind farms over different regions. Our results indicate that by distributing wind farms over only one to three regions, smoothing results vary considerably depending on the choice of regions. However, as the number of considered regions increases, the smoothing improves, and the particular choice of regions matters less for smoothing effects at the investigated time-scales. This highlights the practical importance of deliberately selecting sites for large-scale wind power production to more effectively smooth the aggregated power.