Fig. 5.

Example of the dynamic construction of a neural network (NN) by the repeated structuring and learning procedure (RS&LP) using the ischemia database (IDB). A total of 147 learning iterations of the scaled-conjugate-gradients algorithm, during which 37 new structures are created, leads from the initial architecture [1] to the final architecture [19 9 9]. The training error decreases monotonously (left panel). Some new architectures (e.g., [12 4 2]) are almost not contributing to a reduction of the training error, while others (e.g., [10 2 1]) strongly decrease the training error. With the introduction of a new architecture, the validation error (right panel) may increase in the first iteration (visible in the figure when the new structures [2] and [10 1] are initialized), but it has to decrease monotonously in following iterations. RS&LP stopped when the validation classification reached 100% correctness, yielding the structure [19 9 9]