When analyzing experimental data the standard procedure is to make various cuts in observed kinematical variables in order to single out desired features. A specific selection of cuts corresponds to a particular set of feature functions in terms of the kinematical variables . This procedure is often not very systematic and quite tedious. Ideally one would like to have an automated optimal choice of the functions , which is exactly what feature recognition ANN aim at. For a feed-forward ANN the following form of is often chosen

which corresponds to the architecture of fig. . Here the ``weights''
and are the parameters to
be fitted to the data distributions and is the non-linear neuron
* activation function*, typically of the form

The bottom layer (input) in fig. corresponds to sensor variables and the top layer to the (output) features (the feature functions ). The hidden layer enables non-linear modeling of the sensor data. Eq. () and fig. are easily generalized to more than one hidden layer.

Using eq. () for the output assumes that the output variables
represent classes and are of binary nature. The same architecture can
be used for real function mapping if are chosen linear, in which
case the outermost **g** is removed from the left hand side of ().

**Figure:** A one hidden layer feed-forward neural network architecture.

The weights and are determined by minimizing an error measure of the fit, e.g. a mean square error

between and the desired feature values (targets) with respect to the weights. In eq. () denotes patterns. For architectures with non-linear hidden nodes no exact procedure exists for minimizing the error and one has to rely on iterative methods, some of which are described below.

Once the weights have been fitted to the data in this way, using labeled
data, the network should be able to model data it has never seen before.
The ability of the network to correctly model such unlabeled data is called
** generalization** performance.

When modeling data it is always crucial for the generalization performance that the number of data points well exceeds the number of parameters (in our case the number of weights ). For a given set of sensor variables this can be accomplished by

- Preprocessing using e.g. Principal Component Analysis.
- Building in
*a priori*known symmetries into the problem -- ``weight sharing''. - Adding complexity terms to the error (eq. ()) to
**regularize**the network. - Inspection of the final network to remove redundant parameters.

Fri Feb 24 11:28:59 MET 1995