Information Technology Reference
In-Depth Information
Fig. 7. Retrieval and generation capabilities.
Solid dots represent the proprioceptive and visual
sampling points of the heavy star-shaped prototype time series (Fig. 4). Dashed (visual) lines and
solid (proprioceptive) lines are the time series generated by the network operated in closed-loop
with 'clamped' PB values as the only input. The PB values have been acquired unsupervised ei-
ther during full training (left) or partial training (right). During partial training (right) the network
has only been trained with the prototype sequences for the light-weight circle and the heavy tri-
angle. Still, the network is able to generate a fairly accurate sensory prediction for the (untrained)
heavy star-shaped object.
Fig. 8. Steps until stable PB values are reached.
Bi-modal sensory sequences for all light-
weight and heavy objects (represented by matching symbols in light and dark gray, respectively)
are consecutively fed into the network. The time courses of PB value 1 (solid line) and PB value
2 (dashed line) during the recognition process are plotted.
retrieved with a very high accuracy. As an example this is shown in Fig. 7 (left) for the
heavy star-shaped object.
The steps needed until stable PB values are reached, which in turn can be used for
recognition, are illustrated in Fig. 8. The bi-modal sensory sequences for all light-weight
