Databases Reference
In-Depth Information
FIGURE 3.5: Collaborative-comparison learning approach for the one-
dimensional pattern “ABCDE.” Each activated graph neuron (GN) stores the
signals received from its adjacent neurons.
The collaborative-comparison learning approach compares an external in-
put pattern to the stored entries of each neuron's bias array, which is a lo-
cal data structure containing the history of adjacent node activation. Each
neuron learns by comparing the signals from its adjacent neighbors and
the recorded entries within its memory, i.e., the bias array. A bias array
σ = {s
1
, s
2
, . . . , s
x
} , comprises signal entries, s
i
for i ∈ x. If the exter-
nal signal set matches any of the stored entries, i.e., s
ext
∈ σ, the bias index,
i, of the matched s
i
will be recalled. Otherwise, the signal will be added into
the memory as s
x+1
. There are two advantages to using this approach: 1) the
bias array design for pattern storage minimizes the data storage requirement
and 2) all types of data can be processed. For instance, the signal can be data
vectors or frequency signals, and thus spatial and temporal data can be accom-
modated. In addition, the collaborative-comparison learning technique does
not require the synaptic plasticity rule used by other learning mechanisms,
such as Hebbian and incremental learning. Thus, new patterns are learned
without affecting previously stored information.
