havBpNet-J Data Overview

havBpNet:J

Intro    Network Classes Overview    Data Classes Overview    Packages

havBpNet:J Home Page

havTData and havTpat data handling classes

Often, when training or consulting a neural network, data has been pre-processed into acceptable patterns for the network’s use. This pre-processing typically consists of collecting the data into <input/expected> pairs and applying various scaling and normalization. Data management inside a network training or consultation application can be a bit tedious in its bookkeeping - randomization of pattern presentation order - ensuring appropriate scaling of input and output data - remembering the scaling parameters used for training so the same can be applied for consultation - etc.

In order to ease the burden of data handling, we have provided two auxiliary classes: havTdata and havTpat. These classes will work for both training and consultation data consisting of <input pattern / expected pattern> pairs or for consultation data with only input patterns.

Since collection of data into acceptable patterns is dependant on the source of the original data (i.e. data may be read from sensors, entered by hand, scanned from documents, etc.) these aux. classes do not provide a first-level collection mechanism. Rather, all data are expected to exist in ascii form in a file on disk. (the expected disk-file format is described in the havTdata member function, ...). The auxiliary classes do provide a convenient and consistent way of handling data within network training or consultation applications.

The havTdata and havTpat classes also implement several scaling mechanisms:

data may be scaled using havSystem.havlog(...),
columns may be normalized using calculated means & standard-deviations,
individual input patterns and/or expected patterns may be normalized using standard vector normalization.

The havTdata class provides a mechanism for collecting and acting on patterns as a group. This class provides management of the collection of patterns and an interface for accessing specific patterns in the set.

The havTpat class provides the instanciation of a single pattern. Each pattern should always have an input-pattern part and may also contain an expected-pattern part.

Data Overview || Data Create || Data Scaling || Data Performance
Data Save & Restore || Data Presentation Order

Intro    Network Classes Overview    Data Classes Overview    Packages