Parameter determination for an implementable feedback neural network
Bo Ling, F.M.A. Salam · 1993 IEEE International Symposium on Circuits and Systems · 2002
The authors describe a method which ensures a designed neural network to be implementable as an electronic circuit. The approach involves two steps: (1) adjust the slope of the sigmoidal function of each neuron based on a given criterion; (2) find the weight matrix by an analytical learning algorithm. It is shown that the slope of the sigmoidal function around the origin plays an important role in the implementable neural network design. Based on the approach, the resistance in the neural circuit can be made very large, which reduces the network power dissipation.>