Information Technology Reference
In-Depth Information
Fig. 5. Error plot comparing classical (left) to modified (right) RNNPB.
The mean squared
error (MSE) of the
sin
sequence is shown in black, whereas the average MSE of the
sinc
sequence is drawn in gray.
Fig. 6. Classification using all object categories for training.
PB values are depicted using
white (light-weight) and black (heavy) symbols matching the corresponding shape. Large sym-
bols represent the class prototypes used for training. Smaller symbols depict PB values obtained
during testing with bi-modal single trial data. If the objects have not been correctly classified,
they are shown gray.
whereas the PB values coding for the triangular-shaped objects are positioned more
distantly. This is due to the deviating visual sensory impression they generate (Fig. 4).
The experiment has been repeated several times with different random initializations of
the network weights. However, the obtained PB values of the different classes always
demonstrate a comparable geometric relation with respect to each other.
To demonstrate the retrieval properties (Sec. 2.2) of the fully trained architecture,
the PB values acquired during training were 'clamped' to the network. Operating the
network in closed-loop mode showed that the input sequences used for training can be
