Increasing returns, learning-by-doing and neural networks