Indeed in both papers the
role of ORing in the complex cell layers as a means of obtaining invariance is emphasized.
However there are major differences between the architecture described
here and the neocognitron paradigm. In our model training is only done for local
features. The global integration of local level information is done by a fixed architecture,
driven by top-down information. Therefore features do not need to get
more and more complex. Namely there is no need for a long sequence of layers.
Robust detection occurs directly in terms of the local feature level which has only
the oriented edge level below it.