Classification of trajectories - Extracting invariants with a neural network

A neural classiier of planar trajectories is presented. There already exist a large variety of classi-ers that are specialized on particular invariants contained in a trajectory classiication task such as position-invariance, rotation-invariance, size-invariance, .... That is, there exist classiiers specialized on recognizing trajectories e.g. independently of their position. The neural classiier presented in this paper is not restricted to certain invariants in a task: The neural network itself extracts the invariants contained in a classiication task by assessing only the trajectories. The trajectories need to be given as a set of points. No additional information must be available for training, which saves the designer from determining the needed invariants by himself. Besides its applicability to real-world problems, such a more general classiier is also cognitively plausible: In assessing trajec-tories for classiication, human beings are able to nd class speciic features, no matter what kinds of invariants they are confronted with. Invariants are easily handled by ignoring unspeciic features.