The inference shifting concept is an historical source of confusion. Please read:
- https://github.com/numenta/nupic/wiki/Online-Prediction-Framework#shifting-inferences - http://lists.numenta.org/pipermail/nupic_lists.numenta.org/2015-September/011579.html --------- Matt Taylor OS Community Flag-Bearer Numenta On Mon, Nov 2, 2015 at 6:29 AM, Wakan Tanka <[email protected]> wrote: > Hello Pascal, > > Yes of course I mean not predictable by NuPIC not anything else ;) I should > say "Can this be generalized that if NuPIC returns bouncing (non stable) > anomaly score TOO OFTEN...". There is always question what "TOO OFTEN" means > but it is for broader discussion. > > No I do not use inference shifter, I was just predicting one step ahead and > when I plot the graph it was obvious what is going on ... I thought > inference shifter is useful with larger prediction steps. > > If I understand correct then without inference shifter prediction made in > time T is for data in time T+1. Anomaly score in time T represent the > "confidence of prediction accuracy" in time T if I can say it like that. > Assume one step prediction for above. > > If I understand you correct then inference shifter allows you to use > "another column" which puts prediction T+1 to T? I did not catch it well > from your last message. Sorry > > Thank you. > > > > > > On 11/02/2015 12:52 PM, Pascal Weinberger wrote: >> >> >> Not predictable using nupic, but as Matt said, give it it's time, or >> better data until you interpret or judge anything :) >> >> But one quick question regarding you first point, if you use the inference >> shifter (as it is done in most tutorials) then the results of the model are >> already shifted to where they belong. So nupic outputs prediction and >> anomaly to (in your case) t+1 but the inference shifter puts them there as >> well... So do you use it? >> >> >> Best, >> >> Pascal Weinberger >> >> ____________________________ >> >> BE THE CHANGE YOU WANT TO SEE IN THE WORLD ... >> >> >>> On 02 Nov 2015, at 10:57, Wakan Tanka <[email protected]> wrote: >>> >>>> On 11/02/2015 06:12 AM, Matthew Taylor wrote: >>>>> >>>>> On Sun, Nov 1, 2015 at 2:26 PM, Wakan Tanka <[email protected]> wrote: >>>>> 1. If this is one step ahead prediction then the prediction value on >>>>> line n should correspond to the original value on line n+1 >>>>> (assuming that NuPIC made good prediction and not mistake)? >>>> >>>> >>>> If the prediction is perfectly right, yes. >>>> >>>>> 2. If first question is true can you please explain me the 179 line? >>>>> On >>>>> line 179 there is prediction which equals 0 and on line 180 >>>>> original >>>>> value equals to 0 which is OK. But why I get anomaly score 1 on >>>>> line >>>>> 179? >>>> >>>> >>>> Just because the best prediction is correct does not mean that the HTM >>>> is confident that it is correct. For example, NuPIC might only be 23% >>>> confident in the best prediction it gives, in which case the anomaly >>>> score could be very high. >>>> >>>>> 3. Or you can look at it vice versa: Prediction on line 180 is equal >>>>> to >>>>> 0 but the original value on line 181 is 3. So I assume prediction >>>>> was wrong. Why anomaly score on line 180 equals to 0? Does it >>>>> means >>>>> that NuPIC believe that it is predicting the correct value but in >>>>> fact it was wrong? >>>> >>>> >>>> I would not pay too much attention to the anomaly score (or >>>> predictions for that matter) until the model has seen a few thousand >>>> rows of data. It looks like it has seen less than 200 rows as this >>>> point, so the anomaly scores can vary wildly until it establishes what >>>> the data patterns are. >>>> >>>> Regards, >>>> --------- >>>> Matt Taylor >>>> OS Community Flag-Bearer >>>> Numenta >>> >>> >>> Hello Matt, >>> >>> Can this be generalized that if NuPIC returns bouncing (non stable) >>> anomaly score then it is either because NuPIC does not see enough data or >>> because the data are not predictable? >>> >>> Thank you very much >>> >> > >
