Staged assimilation: a system for detecting invariant features in temporally coherent visual stimuli

Staged assimilation is a means of extracting information from a sequence of related images. It is used to develop feature detectors of transformation-invariant properties. It is applied within a hierarchical network, referred to as the bipanel architecture. Temporal response properties of neurons are used to capture this information. High-level feature detectors arise when sequentially detected low-level features are grouped into equivalence classes. When the system is exposed to imagery undergoing a continuous transformation, it develops detectors for properties that tend to remain constant over time. This is how it forms invariant feature detectors. The authors explain how the approach derives from studies of the visual cortex and J.J. Gibson´s theory (1986) of perception. An example is given of how it works, and details of the system´s operation are discussed. Its relationship to other learning paradigms and networks is described.<>