Building Linux based neural network applications

Neural networks can be trained to approximate arbitrary nonlinear mappings. Because of this capability, they have been successfully used in applications such as system modeling, time-series prediction, automatic control and pattern recognition. In these applications, a mapping is needed to represent the input-output relationship of a real-world process. Neural networks can be trained to form this mapping. However, process parameters may vary over time. When this occurs, the neural network has to be retrained. If a neural network is already being used in a system, new real-time data has to be collected and used to retrain the neural network. Data collection and retraining have to be conducted without disturbing the main task. The retraining should be automatically initiated when significant errors are detected and should stop when the new neural network is satisfactory. Developing the software for such a neural network based system is not trivial, especially if the application is for embedded systems. The development can be made easier when a multitasking operating system such as Linux is employed. This paper provides the results of the investigation into how such an "adaptive" system can be designed.