New! Bayesian Network
The Bayesian Network is a modified MLP network designed to avoid overfitting. The tendency of neural networks to overfit the training data, resulting in poor generalization, is one of the main areas of concern about their practical use. When the error criterion is the error sum of squares or error mean square, one can always produce a smaller error in the training data by employing a more complex neural network. However, a more complex network might fit idiosyncracies in the training data and might not generalize well. The method of stopped training, where the "best" solution is the iteration in which validation set error is minimized, is a method intended to prevent overfitting and poor generalization. The Bayesian network instead uses a different error criterion that in effect adds an additional term to the error sum of squares. In optimizing this criterion, the neural network produces "smooth" predicted values that tend to generalize better. The Bayesian network does not use stopped training, so you do not need to specify a validation set, and can instead pool these records into the training data set. The Bayesian network tends to train in fewer iterations than the Multi-Layer Perceptron.

New! Data Output Tool 
The new Data Output tool has the same appearance and functionality as the Data Input tool. One can easily pass data through a neural network for scoring and save the results in an SPSS system file or some other popular formats.

New! Confidence Outputs in Radial Basis Function
The Radial Basis Function tool writes a new output variable called Confidence Outputs. When the target variable is numeric and you check the Confidence Outputs check box, the Radial Basis Function generates predicted outputs for the target field and also generates a second output called the confidence output, which is in the same units as the target field and can be viewed as a likely maximum error for that particular output.

New! Model weights available
When a neural network is trained, you can click the neural network tool's Status menu item and view a lot of information about the trained neural network including the model weights. These model weights are specific to the network topology and the way in which the network trained. They can be used to construct a set of equations that embody the neural network. These equations, built outside of Neural Connection, can be used to score new data.

Redesigned! Data Input Tool
The new Data Input Tool has the same form as the Data Output Tool. Now, reading in run data for scoring by a trained neural network is easier. Added functionality includes more information about the variables in the input spreadsheet. The Class Equalization feature is useful when the target variable is a symbolic variable with widely disparate class sizes.