On the training of limited precision multi-layer perceptrons

The effects of quantization on the training dynamics of a real-valued feedforward multilayer neural network when implemented in digital hardware are analyzed. It is shown that special techniques have to be employed to train such networks where all the variables are represented by limited numbers of bits in fixed point format. A training algorithm based on the analysis called the combined search algorithm is proposed. The combined search algorithm consists of two kinds of search techniques and is easy to implement in hardware. Using intracardiac electrograms and sonar reflection pattern recognition, extensive computer simulations were conducted. The simulation results are given.< >