Yeah, each input only sees the last previous input. However, it doesn't see the column-based / SP representation, but the cell-based / TM one. Thus its actually seeing the last input in context of the other previous inputs. My guess is that this could be achieved by including some probability when selecting which cells will represent a context for the fist time, instead of selecting random cells.
The hierarchy certainly would make this feature more useful, though. On Mon, Dec 28, 2015 at 3:36 PM, Chandan Maruthi <[email protected]> wrote: > Sebastine. > My understanding is that the answer is no. As each input sees only the > immediate next input. The representation of D in the 2 cases on TM will be > very different . However with a hierarchy [not yet well tested / well > implemented] this should be achievved. As ABCD will be more similar to ABFD > vs RTHD. > > Chandan > > > On Mon, Dec 28, 2015 at 9:09 AM, Sebastián Narváez <[email protected]> > wrote: > >> Hello Nupic. >> >> I was wondering about the TM mechanisms, and I thought that if you have >> two sequences that are alike (say A-B-C-D and A-B-F-D), and one that is >> "significantly" different (say R-T-H-D), then the representation of "D" for >> the first sequence at the TM stage, should share "significantly" more cells >> with its representation for the second sequence than for the third one. >> Does the current implementation do that? >> >> >> Cheers. >> > > > > -- > Regards > Chandan Maruthi > >
