Databases Reference
In-Depth Information
FIGURE 2.5: Distributed pattern recognition based on the process farming
approach.
sent to the master node after each cycle. This process is iteratively per-
formed until the optimum bias weight and errors have been achieved by
the network. Each processor uses a subset of the training data. There-
fore, each processor performs the training procedure on a subset of the
overall data.
2. Pipelining: The recognition procedure of the pipelining approach follows
an incremental method. The training process is conducted subsequently
using a pipeline procedure similar to that shown in Figure 2.6. Each
processor contains a copy of the algorithm and performs a recognition
process on a particular training subset. However, each time the weight
and error changes are passed from one processor to another, they are
modified, evaluated, and incorporated into each subsequent weight and
error calculation.
The top-down approach towards distributed pattern recognition has several
limitations including the following:
1. Recall Disintegrity: Due to the vertical splitting of data, the distribution
of a training data set into a number of subsets can create disintegrity in
the training process and influence the actual recall process. The weight
changes produced by the algorithm on a highly cohesive training set, i.e.,
