I will describe an approach to the learning and use of multi-level hierarchical representation for the purpose of recognizing categories and individual objects, together with their parts and sub-parts at multiple levels. The acquired object representation uses a hierarchy of shared features, selected by maximizing the information delivered for categorization. The learning process automatically extracts the part-structure of the category, and during recognition, novel members of the category are recognized, and their parts are detected and localized. During learning, features representing different appearances of the same object part are combined in the representation into abstract components. Recognition of objects and their parts is obtained by a feed-forward sweep from low to high levels of the hierarchy, followed by a sweep from the high to low levels. I will discuss some general conclusions from our experience regarding the learning and use of multi-level hierarchies.
BACK
