INTRODUCTION

The large amounts of nannoplankton data collected worldwide require sophisticated and powerful quantitative methods of analysis for the extraction of the optimum amount of information on species distributions through which paleoceanographic reconstructions might be based. One of the analytical methods now available is the artificial neural network (ANN) approach for data analysis and forecasting.

Artificial neural networks are computer systems that have the ability to `learn' a set of output, or target, variables from a set of input variables. Neural networks have been employed in many disciplines for problems of prediction, classification, or control of various processes. This remarkable success can be attributed to a few key factors.

• Neural networks are numerically sophisticated modelling techniques, capable of modelling extremely complex functions. For many years, linear modelling has been the most commonly used modelling technique in paleo-oceanography because linear models have well-known optimization strategies. In cases where the linear approximation was not valid (which was frequently the case in nature) the models suffered accordingly. Neural networks analyse the dimensionality problem. This distinguishes them from attempts at modelling nonlinear functions with large numbers of variables;
• Neural networks learn by example. The neural network user gathers representative data, and then invokes training algorithms to automatically learn the structure of the data. In addition, neural computers have the ability to learn from experience in order to improve their performance, and can adapt their behaviour to new and changing environments. The level of user knowledge needed to successfully apply neural networks is much lower than would be the case using more traditional statistical methods;
• Neural networks tend to be more reliable and versatile than conventional data analysis and modelling methodologies. They have the ability to cope well with incomplete (e.g., small samples), `fuzzy' data, and can deal with previously unspecified or unencountered situations. Neural networks are very tolerant to analytic faults. This contrasts with conventional systems, where the failure of one component of the analysis usually means the failure of the entire analytic system.

In the earth sciences, neural networks have been applied to problems of well-log interpretation (Baldwin et al. 1989; Baldwin et al. 1990; Rogers et al. 1992), for the identification of linear features in satellite imagery (Penn et al. 1993); for geophysical inversion problems (Raiche 1991), for the correlation of volcanic ash layers (Malmgren and Nordlund 1996); and for the establishment of present-day climatic zonation in Puerto Rico (Malmgren and Winter 1999). Malmgren and Nordlund (1997) applied a BP neural network in an attempt to predict modern sea surface-water temperatures (SST) from relative abundances of planktonic foraminifer species in the southern Indian Ocean. That study showed the BP technique to be able to reproduce the SST data more faithfully than conventional techniques such as the Imbrie-Kipp Transfer Functions (Imbrie and Kipp 1971) and Modern Analog Technique (Hutson 1979). These results indicated that late Quaternary summer and winter SST's may be predicted with a precision of ±0.7-0.8ºC using the trained BP network.

With data gleaned from the literature we here make a first attempt at testing the applicability of ANN to reconstruct SST and stable isotope data from relative abundance data of calcareous nannoplankton species.