Databases Reference
In-Depth Information
counterpart. After the message communication between adjacent
neurons is completed, the active neurons will update their bias ar-
rays and send the stored/recalled index/indices to the neuron at
the same position in the higher layer (except for the neurons at the
edges). This stage will be repeated for each layer above the base
layer, until the top layer neurons are reached.
Stage 4.
One of the top layer neurons will receive a bias index from a neuron
in the layer underneath. This top layer activated neuron node will
search its bias array for the index. If the index is found, this node
will trigger a recall flag with the recalled index. Otherwise, it will
trigger a store flag and store the new index in its bias array. It will
send a signal message to the SI module with the message format
{sn
id
, sn
st
, sn
idx
}, where status is either recall or store. The signal
message sent by the top layer active neuron marks the completion
of the recognition procedure at the subpattern level. In a DHGN
implementation, the lower bias arrays are updated when a new entry
is found. Note that the bias index for lower layer neurons might not
be the same for a given pattern index.
Figure 5.3 shows the process workflow of the proposed recognition algorithm.
5.1.3.2
Phase 2: Pattern Reconstruction and Recognition
Recognition results obtained by the SI module from all subnets in a DHGN
network require further analysis to derive an overall recall of the respec-
tive input subpattern. Two methods have been considered. These are recall-
percentage and voting methods. These methods differ in terms of the mech-
anism adopted. This research intends to compare and contrast these two ap-
proaches from an accuracy perspective.
5.1.3.2.1 Recall-percentage method
The recall-percentage method
underlines the use of bias indices obtained from all neurons in each subnet.
The main principle of this approach is that the recall/store decision is based
on the cumulative decisions of all neurons in the network.
This method requires that an additional procedure be conducted by
each DHGN subnet for the purposes of index collection before final
recognition results are submitted to the SI module. For each subpat-
tern introduced into the subnet and after all of the recognition pro-
cesses have been completed, the activated top neuron will collect all of
the index information (idx) from all of the neurons underneath it. These
indices are be compiled and structured with the format {idx, count}.
These outputs are sent to the SI module using the message format
{sn
id
; (idx
1
: count
1
) , (idx
2
: count
1
) , (idx
3
: count
1
) , . . . , (idx
n
: count
n
)}for
all n indices recalled or stored.
