In the last two decades significant progress has been achieved in theoretical aspects of arificial neural networks. Software simulations based on the theory of artificial network paradigms are currently available in conventional (serial) computers. This often constitutes a bottleneck in practical applications, where tranining (teaching) time can take days or weeks. Software simulations of artificial neural networks are much slower in comparison with signal processing by circuits (hardware). Therefore, the practical use of neural networks is heavily based on circuit implementations.
Synthetic neural networks can be implemented in silicon either digital or analog integrated circuits, or in a analog/digital mixed form. The largest computational load in a neural system is incurred by the weighted summation $\SigmaT_{ij}$·$V_j$, where $V_j$ is a neural state and $T_{ij}$ the matrix of synaptic weights. Hardware implementations of neural networks have a host of approaches. Among the approaches, pulse coding neural networks analogous to real biological neural networks are most promising in that those have the merits of both analog and digital technology. But, it is a disadvantage that the pulse coding approaches need more time than the analog ones and the binary coded digital ones to compute the weighted summation. Also, the exact speed calulation method of the pulse coding neural networks is not established.
In this thesis, I will propose three novel neural circuits for pulse coding neural network implementation. The performances of the proposed neural circuits are compared, using a novel method to measure the computation speed of the pulse coding neural circuits. The artificial neural circuit using noise feedback pulse coding(NFPC) needs far fewer pulses than that using stochastic pulse coding(SPC) or integration-fire/reset pulse coding(IFRPC) to convert an analog neuron body voltage into a pulse sequence under the same signal-to-noise ratio. This result suggests that the NFPC neural circuit can improve the speed of computation and communication, compared with SPC one and the IFRPC one. In order to represent the excitatory and inhibitory synaptic inputs to the neuron body, I use a simple transconductor circuit. Also, I will present a novel concept and evidences that real biological neurons use a kind of NFPC, and propose a novel neuron model based on the concept. Simulation and experimental results are included.
