Each model cooperates with the image depending on its similarity. The most similar model cooperates most successfully and is the most active one. Hence, the total activity of the model layers indicates which is the correct one. We have derived a winner-take-all mechanism from [4] evolution equation and applied it to detect the best model and suppress all others. The corresponding equations are (cf. Equations 1 and 7):
The total layer activity is considered as a
fitness , different for each model
. The modified evolution equation can be easily analyzed if the
are assumed to be constant in time and the recognition
variables
are initialized to 1. For the model layer
with the highest fitness, the equation simplifies to
with a stable fixed point at
. For all
other models the equation then simplifies to
, which results in an
exponential decay of the
for all
. When a
recognition variable
drops below the suppression threshold
, the activity on layer
is suppressed by the term
. The time scale of the
recognition dynamics can be controlled by
.