We will present a model in which meaningful understanding of scenes is obtained  from the combined processing of a bottom-up (BU) and top-down (TD) streams, interacting through a bi-directional communication between them. The BU stream creates a partial visual representation in the higher level parts of the model. The model then provides a top-down instruction to the TD stream, which guides the next cycle, to extract selected information and expand the existing representation.  By automatically selecting an appropriate sequence of TD instructions, the model successively extracts from the scene structures of interest in a goal-directed manner. We will show recent results of extracting complex scene structures, and the ability of the model to reach broad generalization to novel scene configurations.

Zoom link: https://mit.zoom.us/j/93324663073