Towards a more powerful, multi-linear
stackable learning module
So far, the states of the units in one layer have only been
used to determine the effective biases of the units in the
layer below.
It would be much more powerful to modulate the pair-wise
interactions in the layer below. (A good way to design a
hierarchical system is to allow each level to determine the
objective function of the level below.)
For example, a vertical edge represents a breakdown in
the usual correlational structure of the pixels: Horizontal
interpolation does not work, so it needs to be turned off,
but interpolation parallel to the edge is OK.
To modulate pair-wise interactions we need higher-order
Boltzmann machines.