Databases Reference
In-Depth Information
FIGURE 4.4: Total number of neurons in an HGN for one- and two-
dimensional compositions as a function of pattern size.
pair to all of the adjacent neurons. For each active neuron in the base layer,
the p(column, row) pairs received from adjacent neurons make up the bias
array entry for the current input pattern. In the end, each non-edge neuron
received two pairs from its adjacent neurons; neurons on the edges receive a
single pair. Each active neuron must determine its bias index. If the incoming
pair combination is found in its bias array, then the index of the entry is
noted. Otherwise, a new index is generated to store and reference the pattern.
Each active non-edge neuron sends its index to its corresponding neuron in
the same column of the higher layer. This process continues until the top layer
has been reached. The top layer neuron decides if the input is to be treated
as a new pattern and stored or treated as a previously known pattern and
recalled. A new index value is propagated downward for a stored pattern, and
an existing index value is propagated downward for a recalled pattern.
In the HGN recognition procedure, the bias array structure of the hierar-
chical composition follows the bias array formation in a GN network. Nev-
ertheless, it has been modified to accommodate the recognition procedures
of higher layer neurons based on adjacency comparisons made by lower layer
neurons. These are the bias entry conditions for neurons within any HGN
network:
