Prevailing computational models of recognition are purely feed-forward. In these
models, as sensory input passes from region to region more abstract features are
extracted until several levels up in the hierarchy entire objects are recognized.
This passive view is inconsistent with anatomical and physiological experiments,
which suggest that active movement is integral to every cortical region, including
primary and secondary sensory areas. Integrating motion and feed-forward sensory
inputs into a unified model of recognition remains an important open problem.
We describe a network model of cortical circuitry that learns sensorimotor
representations of objects. Extending previous work, the network integrates motor
representations and feed-forward sensory information to build predictive models
of objects. Objects are actively explored and then recognized quickly as soon as
sufficient disambiguating information has been sensed.
The model explains a number of empirical observations that have eluded theoretical
understanding, including the role of bidirectional connections between “what” and
“where” regions, some of the inter-layer and intra-layer connectivity patterns
within cortical columns, and why some layers of cells have extensive lateral
connections between columns.