| |
Abstract:
Abstract: A growing body of evidence indicates that object
recognition is slower the further a test object is rotated from its
studied or canonical orientation. Such results are often ascribed
to the operation of a mental rotation-like process that brings
novel views of an object into register with encoded templates, but
the implementation of such a process in a working model is
problematic. We present a simple model that recognizes objects on
the basis of templates without the need for viewpoint
normalization. The model's input representation is passed through a
set of connections to an abstract representation layer (ARL), which
in turn connects to a response layer where "evidence" for various
recognition responses builds up. Over the course of a trial, the
model shifts the image around so that the input representation is
centered on different parts of the object. The model generalizes
very well to small viewpoint changes and performs less well, but
far better than chance, on many larger viewpoint changes. It also
generalizes to untrained exemplars of object categories. It can
learn common objects, novel objects, letters, and faces, and can be
taught to recognize a large number of any combination of these
stimulus types at the same time. Individual ARL units show response
functions similar to those of neurons in monkey inferotemporal
cortex. These successes, as well as limitations to the model, will
be discussed.
|
