Bias-variance dilemma (Geman et al., 1992). It can be demonstrated that the mean square value of the estimation error between the function to be modelled and the neural network consists of the sum of the (squared) bias and variance. With a neural network using a training set of fixed size, a small bias can only be achieved with a large variance (Haykin, 1994). This dilemma can be circumvented if the training set is made very large, but if the total amount of data is limited, this may not be possible.
Latest Updates: Silvert 1998 RSS
Perhaps the greatest problem that is faced in most attempts to use artificial neural networks for ecological applications is that the quantity of data is often very limited. Although there are a few cases where large amounts of data are available, as in the case of remote sensing or observations based on automatic telemetry, it is far more common to have to deal with limited and irregularly spaced data, and the data may not always be strictly comparable due to variations in environmental conditions between sampling periods. In most situation the collection of field data is both time-consuming and expensive.
Since the training and testing of neural networks is very data-intensive, this poses serious obstacles to the development of neural network applications in ecology.
Silvert W, Baptist M. 1998. Can neuronal networks be used in data-poor situations? Di dalam: Lek S, Guégan JF. Artificial Neuronal Networks: Application to Ecology and Evolution. Berlin: Springer-Verlag. hlm 241-248.