A DSOM hierarchical model for re exive processing : anapplication to visual trajectory classi

Any intelligent system, whether human or robotic, must be capable of dealing with patterns over time. Temporal pattern processing can be achieved if the system has a short-term memory capacity (STM) so that diierent representations can be maintained for some time. In this work we propose a neural model wherein STM is realized by leaky integrators in a self-organizing system. The model exhibits compositionality, that is, it has the ability to extract and construct progressively complex and structured associations in an hierarchical manner, starting with basic and primitive (temporal) elements. An important feature of the proposed model is the use of temporal correlations to express dynamic bindings.