Next: A. Statistical tools Up: 8. Conclusions Previous: 8.5 Perception

# 8.6 Future direction

The present work can be extended in several directions:

• To remove the sensitivity of NGPCA and NGPCA-constV on the parameters, an automatic adjustment during the training would be helpful. A further helpful extension would be a growing mixture model that adjusts the number of units m and principal components q to the data distribution. For the Neural Gas vector quantizer, Fritzke (1995) developed a model that adjusts m. Meinicke and Ritter (2001) extended MPPCA to adjust q.

• Moreover, for NGPCA, different error measures or different ranking functions might avoid `dead units' and thin ellipsoids that protrude out of a distribution (see section 3.3.2, figure 3.8 and the discussion in section 3.5).

• The recall in the abstract RNN has discontinuities (for more than one unit). An interpolation between neighboring units might lead to better results.

• So far, the robot arm did only grasp objects. A possible extension is to include object manipulation. By mentally simulating such a manipulation, the robot could recognize the object, as in the following example. Two objects need to be recognized: a cylinder and a brick (as in chapter 6). Both lie on the table. If the robot pushes the cylinder, it will roll; if the robot pushes the brick, it will move only a short distance. Once trained, the robot sees an object with its camera. Using simulation, the robot can predict what would happen to the object if pushed. Based on the outcome, the robot can decide if it was a cylinder or a brick. If this experiment also works for different illuminations and different object orientations, it could show that the sensorimotor approach can explain object constancy.

• The mobile-robot experiment may be extended to the perception of dead-ends, as suggested by Möller (1999). Standing in front of a potential dead-end, the robot simulates the outcome of an obstacle-avoidance algorithm. If the robot predicts that it will get stuck, it can conclude that it faces a dead-end. With the same mechanism, dead-ends of different shape and seen from different perspectives could be perceived (object constancy).

• The abstract RNN could be helpful for other applications. It could be applied to any pattern-association with locally continuous mappings between patterns.

Next: A. Statistical tools Up: 8. Conclusions Previous: 8.5 Perception
Heiko Hoffmann
2005-03-22