Using constructive learning in embedded systems engineering

Embedded systems differ from many other engineering applications in two essential requirements: they are usually restricted to use slow processors, and they must fit within a reduced amount of memory. One of the main claims within the neural networks field is that once trained they are very fast to process. However, many neural network structures need a respectable amount of memory to maintain their information. This paper shows how constructive learning methods can be used to gradually increase a feedforward neural network complexity to achieve an optimal trade-off between the desired training error and memory requirements. This is a very important issue in engineering design tasks and applications, especially for embedded systems. In addition, the constructive training method is reviewed, a practical application addressed and the results obtained discussed.