|
Results:
-
Extension of the Modeling Field
Theory neural network for the classification of objects (as
seen in ICANN 06)
the system is able to dynamically
adapt when an additional feature is introduced during learning
Description:
-
Test 1:
Consider the case in which we have
the 5 objects, initially with only one-feature information. We
can consider color information only on Red, the first of the 3
RGB feature values,
The objects have the
following R feature values: O 1
= [0.1], O2 = [0.2], O3 = [0.3], O4 = [0.5], O5 = [0.5].
Click for full size image
Figure 1: Time evolution of
the fields with only the first feature being used as input. Only
4models are found, with two initial random fields converging
towards the same .5 Red concept model value.
-
Test 2: Let us
now consider the case in which we add information from the
second color sensor, Green. The object input data will now look
like these: O1 = [0.1, 0.4], O2
=[0.2, 0.5], O3 = [0.3,
0.2], O4 = [0.5, 0.3], O5
= [0.5, 0.1]. The same MFT algorithm is applied
with 5 initial random fields. For the first 12500 training
cycles (half of the previous training time), only the first
feature is utilized. At timestep 12500, both features are
considered when computing the fuzzy similarities.
Click for full size image
Figure 2: Time evolution of
the fields when the second feature is added at timestep 12500.
Thedynamic fuzziness reduction for s2 starts at the moment the
2nd feature is introduced, and is independent from s1.
Up -
Back
|
