Classification and position adjustment is carried out via a full connection layer (FC), a data set before an after full connection layer parameters , dimensions , once the forward propagation (Forward) namely:
Computational complexity .
It will be the SVD, with a front and a value of approximately t features:
The original exploded front propagating into two steps:
Computational complexity becomes .
When implemented, equivalent to a split into two fully connected layers, are connected to a middle lower dimension: