The early Numenta information hinted that an HTM machine can be similar to this: it takes data and generalise it and then next layer will take the generalised data and generalise it more. so in a sense 1st layer would recognise some elementary patterns, 2nd - patterns of those patterns and so on…
There are quite a few things changed with HTM perception at Numenta in past 4 years…
Can I still do this? Can it be done using Nupic? Can I predict, using unfinished sequence, proabilities of patterns this sequence might form…
like 112 can form a previously recognised (fitted|trained) patterns: 11223 or 11221122 or 112112…
so as the result of predicion on each time step to have {x0, y0}, {x1,y1}… where x is pattern (11223 or 11221122 etc) and y is probability of 112 finishing as x.