Hitoaki Yoshida

Origin of Randomness on Chaos Neural Network

We have proposed a hypothesis on the origin of randomness in the chaos time series of a chaos neural network (CNN) according to empirical results. An improved pseudo-random number generator (PRNG) has been proposed on the basis of the hypothesis and contamination mechanisms. PRNG has been implemented also with the fixed-point arithmetic (Q5.26). The result is expected to apply to embedded systems; for example the application of protecting personal information in smartphone and other mobile devices.

KANSEI (Emotional) Information Classifications of Music Scores Using Self Organizing Map

We classified KANSEI (emotional) information for musical compositions by using only the notes in the music score. This is in contrast to the classification of music by using audio files, which are taken from a performance with the emotional information processed by the instrumentalists. The first is classification into one of two classes, duple meter or irregular meter. The second is classification into one of the two classes, slow vs. fast (threshold tempo: ♩ = 110). The classification of the musical meter is based on identifying the meter indicated in the score. For tempo classification, we generally used the tempo indication in the score, but we evaluate classification that includes tempo revisions through a subject’s emotions to be accurate. We performed classification for both the meter and tempo evaluations with a recognition rate above 70 % by using self-organizing maps for unsupervised online training. Particularly, in the tempo classification, a computer successfully processed the emotional information directed.

FPGA Implementation of Neuron Model Using Piecewise Nonlinear Function on Double-Precision Floating-Point Format

The artificial neurons model has been implemented in a field programmable gate array (FPGA). The neuron model can be applied to learning, training of neural networks; all data types are 64 bits, and first and second-order functions is employed to approximate the sigmoid function. The constant values of the model are tuned to provide a sigmoid-like approximate function which is both continuous and continuously differentiable. All data types of the neuron are corresponding to double precision in C language. The neuron implementation is expressed in 48-stage pipeline. Assessment with an Altera Cyclone IV predicts an operating speed of 85 MHz. Simulation of 4 neurons neural network on FPGA obtained chaotic behavior. An FPGA output chaos influenced by calculation precision and characteristics of the output function. The circuit is the estimation that above 1,000 neurons can implement in Altera Cyclone IV. It shows the effectiveness of this FPGA model to have obtained the chaotic behavior where nonlinearity infuences greatly. Therefore, this model shows wide applied possibility.

Implementation of simple neural network system based on pulse-train arithmetic

In this paper, implementation and operation of a simplified Hopfield type neural network system based on pulse-train arithmetic are presented. The narrow module proposed here as a basic building block is composed of a pair of neuron unit and synapse unit, which performs time-division multiple operation. The Hopfield type neural network is implemented by connecting the neuron modules in parallel with a single bus line, so that this structure can considerably reduce a mass of hardware. In particular, the network can easily be extended to more practical systems from the simple structure.<<ETX>>