Multi-vector sampling prevent overlap #3
Description
Currently when sampling multiple vectors superposed, we don't care whether any of these vectors are exactly the same, i.e. composed of the same factors.
Depending on the workload we are modeling, we may need to disallow overlapping to happen.
The resonator network algorithms do not perform extra check when the same vector is extracted multiple times, even if overlap is not allowed - it will just lead to a (likely) wrong answer.
Another phenomenon is that when a bundled vector is composed from multiple instances of a majority vector plus some other single-instance minority vectors, the resultant vector is pulled toward the majority vector and become almost orthogonal to the minority vectors. This affects the ability to extract the minority vectors unless one or more instances of the majority vector is explained away first.
This presents an issue where, when the count is unknown, we must compare the similarity between the extracted vector and the remaining bundled vector (after explain-away) and the similarity threshold to determine whether it is a valid vector.