Vectors

which consist of the output of every neuron in same hidden layer corresponding to different samples

should be nonlinearly correlated. With this basic fact

this Paper firstly gives the definition of linearly correlated vector and corresponding nonlinear correlation measure for every hidden layer

then adds or deletes a neuron for a hidden layer according to its nonlinear correlation measure and adjusts the neural networks weight

values appropriately. This method can not only avoid confining the number of neuron units in a hidden layer

but also escape local minimum during the learnig process. According to error analysis in detail

if gives the optimistic rule of deleting neuron in hidden layer. Numerical experiments illustrate its efficiency.