A neural network learning algorithm based on linear regression that determines and uses target values for hidden neurons is proposed. An additional key component of the proposed algorithm is the use of linear regression weighting factors derived from a physical representation of the neural network with spring representing weights. The algorithm applies linear regression to each neuron and requires very few iterations. It appears to have significant efficiency advantages over backpropagation for some situations, and, unlike a one-step linear approach, it is capable of learning nonlinear relationships.
展开▼