Study on Flexible Neural Network for Olfactory Sensory System

The authors introduce the flexible stucture of neural network and its algorithm for olfactory sensory system. This method includes two steps: at first, we built a lager network structure with more hidden nodes which is completed off-line according test gases requirements, and then by on-line teaming and calculating the correlativities between hidden neurons and the dispersivities of ones, useless or low-active -neurons are merged or deleted to get feasible smaller network. Besides, according to gase environments can adaptively choice the transfer characteristics of each neuron, that is activation functions to get an optimum function. The experimental results show the above mentioned methods can reduce the structure of network considerably, increase its flexibility or adaptivity, while recognition accuracy of mixed gases is improved too.