The existing VLSI neural network based on pulse width modulation (PWM) technique is analyzed and its new building blocks are proposed. A simple synapse multiplier with high precision and large linearity range is designed, which has no switching noise effects. To transform the neuron voltage state to a PWM signal, a voltage-pulse conversion circuit with high conversion precision and linearity is suggested. To verify the building blocks, a 2-2-1 PWM VLSI neural network is designed, its backpropagation (BP) learning algorithm is adjusted according to the circuit characteristics. The simulation result shows its ability to solve AND, OR and XOR problems. Its speed is more than 1000 times faster than software simulation.
[1]
Bingxue Shi,et al.
Circuit design of an adjustable neuron activation function and its derivative
,
2000
.
[2]
Tetsuro Itakura,et al.
Neuro chips with on-chip back-propagation and/or Hebbian learning
,
1992
.
[3]
Chung-Yu Wu,et al.
Realization of the CMOS pulsewidth-modulation (PWM) neural network with on-chip learning
,
1998
.
[4]
T. A. York,et al.
Hardware implementation of a pulse-stream neural network
,
1998
.
[5]
Ezz I. El-Masry,et al.
Implementations of artificial neural networks using current-mode pulse width modulation technique
,
1997,
IEEE Trans. Neural Networks.