Databases Reference
In-Depth Information
FIGURE 5.2: Analogical representation of the DHGN distributed pattern
recognition scheme.
messages from their adjacent neighbors will respond by updating
their bias array, noting the activation signals. All other neurons
will remain inactive.
Stage 2.
All active neurons in the base layer will update their bias arrays. If
the bias entry value, σ
ent
(left, right) received from the activated
neurons in both the preceding and succeeding columns have been
recorded, the index of the entry is sent to the respective neuron in
the same position at the higher layer. If the σ
ent
(left, right) value
is not found within the bias array, a new index will be created and
sent to the neuron in the higher layer. Note that active neurons
at the edges of the base layer do not communicate with higher
layer neurons. Because of the pyramid-like structure of the DHGN
subnets, there are no neurons present at the edges of a higher layer.
Stage 3.
The DHGN neurons in the layer above the base that receives a
signal message, containing the index of the bias entry that has
been created or recalled from stage 2, will be activated. A pro-
cess similar to stages 1 and 2 will occur. However, the contents
of the signal messages from preceding and succeeding columns will
be in the form σ
ent
(left, middle, right) for non-edge neurons and
either σ
ent
(left, middle) or σ
ent
(middle, right) for the edge neu-
rons. The values for left, middle, and right are derived from the
indices retrieved from the lower layer neurons. For instance, left
is for the preceding neuron's index received from its lower layer
