Modeling and characterizing novel pulsed neural networks for real-time applications
Date of Issue2011
School of Electrical and Electronic Engineering
Biologically inspired artificial neural networks have been well researched and found in many applications. In contrast to the conventional artificial neural network models, pulsed neural models can explain some phenomena of information coding with more realistic biological interpretation. In this thesis, we propose and investigate some new models of pulsed neural networks, including pulse-based radial basis function (RBF) network and multi-channel pulse coupled neural network (MPCNN). The proposed neural computation is suitable for full parallelism implementation, which is a salient merit that can be exploited for a multitude of real-time applications. Different from conventional RBF networks, the proposed pulse-based RBF network utilizes the synchronism of temporal pulses generated by model neurons to perform RBF computations. This bio-inspired design is highly advantageous due to its parallelism characteristics. Based on a novel winner neuron selection algorithm, which is designed by employing the rank order of neurons’ output pulses, it gives comparable performance when benchmarked with the conventional counterparts in clustering and nonlinear function approximation applications. The rank order based paradigm is also extended to deal with some complex clustering problems such as processing of meshed clusters which is difficult if conventional methods such as k-means and self-organizing map (SOM) are used. As a 2-dimensional (2-D) variant of pulsed neural networks, the proposed MPCNN overcomes the shortcoming of the conventional scalar-based PCNN for processing the multi-spectral images. MPCNN is a multi-spectral image segmentation approach that uses the timing of individual pulse produced by the neurons to determine the distances between pixels’ feature vectors and their rank order respectively. Consequently, the fast links can be established among neurons with respect to the spectral feature vectors and spatial proximity of mapped pixels.
DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Control engineering