Data analysis specialist SPSS Inc has extended its neural network software with the release of Neural Connection 2.0 to enable users to do ‘what if ?’ analysis of data, and to learn more about the data model produced by the neural network. Neural Connection is a personal computer-based neural network analysis package developed by Recognition Systems and marketed exclusively by SPSS (CI No 2,739). Version 2.0 includes what the company calls Bayesian network technology, based on the Bayesian philosophy of calculating the unknown, and enables analysis of data without having a full set of validation data. Traditionally, neural networks, which are capable of ‘learning’ patterns in data presented to them, need a complete set of validated data from which to learn. Nodes are given various weightings, and an answer is produced. The trainer then tells the network how far out the answer is to the answer expected from the validated data, and the nodes are re-weighted until eventually the answers begin to tally. The network can then begin to analyze its own data without further programming. Neural Connection Version 2.0 enables users to print displays of the weighting applied to nodes and parameters, so that they can understand more about the final model, and use it in further analysis. This goes towards answering criticisms of neural networking as a data analysis tool, where it has been said that the neural network is a ‘black box’ that the statistician cannot get at to find out how it arrived at its answers (CI No 2,736). Version 2.0 runs the Bayesian analysis alongside more traditional statistical analysis techniques such as Multi-Layer Perceptron, Radial Basis Function and Kahonen Analysis. SPSS says Neural Connection 2.0 provides the ability to explore complex models and predict the most likely outcomes. The models can be adapted to simulate different scenarios by adjusting the parameter and weight of each factor. The product is aimed at five key markets, market research, medical research, financial research, database marketing and scientific research. The software is available now and runs on a any personal computer from an 80386 and upwards, on Windows 3.1 or later. The only extra hardware specification is a mathematical co-processor which is strongly recommended to run the software.
á
