Sensorimotor integration using an information gain strategy in application to object recognition tasks T Kluth1, D Nakath2, T Reineking2, C Zetzsche2, K Schill2 |
|---|
1University of Bremen, Germany |
| Humans can recognize 3D objects robustly and accurately. There is evidence that in natural settings this competence involves not only sensory processing but also motor components. This is not only true for the recognition act itself but also for the representation. However, while we have powerful models for pure sensory processing (hierarchical feed-forward networks), models for a sensorimotor approach to object recognition are rare, and do often address only part of the problems. In particular, it is not yet clear what the specific relations between motor states and sensor information are, and how they enter into the underlying representation. Here we developed and implemented a probabilistic model for object recognition which combines motor states and bottom-up processes of feature extraction in an integrated sensorimotor architecture. The top-down process computing the next movement of the robot is modeled by an information gain strategy which uses a sensorimotor knowledge base to obtain the most informative motor action. In a training phase the knowledge base is learned from real data to obtain the sensorimotor representation. We show how the integration of motor actions effects task performance in comparison to the modeling approach which only takes sensor information into account. |
| Up Home |
|---|