On memory size occupied by a multilayer perceptron

When a multilayer perceptron (MLP) is trained by the backpropagation (BP) algorithm, it is difficult to estimate Q₁, where Q₁ denotes the memory size occupied by the weights and thresholds in the MLP, because of unpredictability of the number of hidden layers and the number of neurons in each hidden layer. By introducing a training difficulty factor, the paper overcomes the difficulty in estimation and gives Q₁. After comparing Q₁ with Q₂, where Q₂ denotes the memory size occupied by the pattern pairs themselves, the result is that only real pattern pairs can be compressed by MLPs.