Using history to predict the future is a necessary skill of intelligence being. We could map the current sequence of events with history to predict the possible future. Then we remember it to accumulate more examples for future prediction.
However, sometimes the mapping is unnecessary between the same objects. In this example, the pattern is that “p1->p2,p2->p1.” Given a->b, complete the rest. If “a” is mapping to “p1” and “b” is mapping to “p2,” we could fill the unknown parts with this relation (“a->b, b->a”).
So I want to ask are there researches in related areas?
Generating patterns with a small amount of data. In this example, given examples 1 and 2, there is an assumption of “p1->p2, p2->p1.”
Discovering mapping between object and pattern.
Completing the unknown parts with a given mapping and pattern.
A big part of the answer is temporal memory (active dendrites). But we also need to incorporate movement (sensorimotor integration) and the warping of space and time (thalamus?). All these questions are pretty central to the ongoing theory work we do at Numenta. Watch our research meetings to follow closely.
