Under Construction.
Let me describe what I mean by learning lower level patterns in the context of higher level patterns. Consider figure ***. Let that box correspond to a V1 region as in figure **. Input comes in at the bottom of the box. Lets denote the inputs by random variable Z.
Now assume that there are some high level concepts that we denote by the random variable Y. We will find out later how these concepts can be formed. Now lets say we have a high level conect Y=y1 as active while the input patterns keep changing. We can learn which input patterns Xs occur frequently while the high level context Y is active by just counting the frequency of occurences of Xs whenever Y is active. In other words, if we know what Y is and when Y is active, we can learn the conditional probability distribution (actually, a conditional probability mass function) P(X|Y). This is what I mean by learning Xs in the context of Ys.