@Yuwei @cogmission Thank you for your inputs. @Yuwei, I haven't experimented high-order sequences yet, But it seems like that's the next place to experiment.
On Fri, Aug 7, 2015 at 1:54 PM, Yuwei Cui <[email protected]> wrote: > @Chandan > In your example sequence 1, it seems that "AAABXY" is repeatedly followed > by itself. It will then be treated as a high-order sequence. So the second > time it sees B in sequence 1, it will only predict "X". > > However, if the subsequence "AAABXY" and "AAABCD" are randomly > interleaved, temporal memory won't be able to learn the "AAABXYAAABXY" as a > high-order sequence. I think it will predict both "C" and "X" after then > 2nd B in that scenario. > > Yuwei > > On Fri, Aug 7, 2015 at 12:28 PM, cogmission (David Ray) < > [email protected]> wrote: > >> Its true that after repeated submissions of the two sequences, the >> Classifier will vote on X or C's bucket with more reliability. Otherwise, >> from what I understand, the TemporalMemory will look for active segments >> leading from its active cells (cells in the column(s) indicating "B"), to >> see which Segments have Synapses who's permanence is above minThreshold, >> and those will be the "predicted" Synapses; and those post-synaptic cells >> will be the predicted cells - which belong to columns indicating the >> TemporalMemory's next prediction after "B". >> >> How's that for confusion? :-) >> >> >> >> On Fri, Aug 7, 2015 at 2:00 PM, Chandan Maruthi < >> [email protected]> wrote: >> >>> @cogmission >>> If thats right i get it, but it doesnt make sense at the the 2nd B you >>> should know that there is a high probabilty of X or C based on the most >>> recent context >>> >>> >>> >>> >>> On Friday, August 7, 2015, cogmission (David Ray) < >>> [email protected]> wrote: >>> >>>> Hi Chandan, >>>> >>>> He's saying that nothing determinant can be predicted at B - and all >>>> possible sequences that are equally predictable will therefore be predicted >>>> because at B, both sequences are ambiguous or equally probable. >>>> >>>> Does that help? >>>> >>>> >>>> On Fri, Aug 7, 2015 at 1:10 PM, Chandan Maruthi < >>>> [email protected]> wrote: >>>> >>>>> Yuwei, >>>>> So you you are saying that at the 2nd B it should be able predict if >>>>> its in the X or C sequence is that right? How does this work? >>>>> >>>>> >>>>> On Friday, August 7, 2015, Yuwei Cui <[email protected]> wrote: >>>>> >>>>>> Hi Chandan, >>>>>> >>>>>> It is not possible to disambiguate the two sequences at the >>>>>> highlighted B. So NuPIC will predict both C & X at that point. However, >>>>>> only one of the predictions will be confirmed at the next step. So if we >>>>>> are indeed in sequence 1, it will predict only Y after X, and vice versa. >>>>>> >>>>>> In other words, TM handles branching temporal sequences by >>>>>> maintaining predictions about multiple possible inputs until there is >>>>>> sufficient disambiguating evidence. Does it make sense? >>>>>> >>>>>> Yuwei >>>>>> >>>>>> On Fri, Aug 7, 2015 at 10:09 AM, Chandan Maruthi < >>>>>> [email protected]> wrote: >>>>>> >>>>>>> Question on Synaptic Connections >>>>>>> Consider 2 sequences >>>>>>> >>>>>>> Sequence 1: AAA*BXY*AAA*BXY*AAA*BXY* >>>>>>> Sequence 2: AAA*BCD*AAA*BCD*AAABCD >>>>>>> >>>>>>> >>>>>>> Consider the B highlighted, how does Nupic know that it is in >>>>>>> sequence 1 vs sequence2 >>>>>>> when the transition from A to B happens, how does it know that it is >>>>>>> in the ABX sequence vs ABC. Also once it starts seeing ABX vs ABC, how >>>>>>> does >>>>>>> it know that the ABX sequence is more relavant at the moment.. >>>>>>> >>>>>>> >>>>>>> -- >>>>>>> Regards >>>>>>> Chandan Maruthi >>>>>>> >>>>>>> >>>>>> >>>>> >>>>> -- >>>>> Regards >>>>> Chandan Maruthi >>>>> >>>>> >>>>> >>>> >>>> >>>> -- >>>> *With kind regards,* >>>> >>>> David Ray >>>> Java Solutions Architect >>>> >>>> *Cortical.io <http://cortical.io/>* >>>> Sponsor of: HTM.java <https://github.com/numenta/htm.java> >>>> >>>> [email protected] >>>> http://cortical.io >>>> >>> >>> >>> -- >>> Regards >>> Chandan Maruthi >>> >>> >>> >> >> >> -- >> *With kind regards,* >> >> David Ray >> Java Solutions Architect >> >> *Cortical.io <http://cortical.io/>* >> Sponsor of: HTM.java <https://github.com/numenta/htm.java> >> >> [email protected] >> http://cortical.io >> > > -- Regards Chandan Maruthi
