Hi all, David's summary is spot-on. Let me add a bit more color:
>From an API standpoint, the Network API can fully support hierarchies in the sense that you can chain together regions that send output to higher level regions. Each region is responsible for deciding what should be sent to the next level. The OPF (which is built on top of the Network API) does not really support hierarchy - the OPF currently supports a very restricted set of models such as our streaming anomaly detection and prediction models. You can also bypass the Network API and just chain together low-level algorithm classes yourself. But that is a coding thing and has little to do with algorithms and theory. It's unclear if simply stacking the existing algorithms will do any good (it probably won't). Within our research group we are actively trying to understand and implement hierarchies (and temporal pooling) the same way it is done in cortex. You can see some of our work described here [1]. However we tend to be pretty methodical. We are trying to do things the "right way" and validate algorithms every step of the way. The link below is our third (fourth? fifth?) attempt at hierarchies and we're still not happy with it. We are not even 100% sure exactly what we want from hierarchies - we often revisit this in our internal discussions and Jeff has discussed some his thoughts on the mailing list. I think it would be great if members of the community wanted to do research in this area. I think there is a lot of low-hanging fruit here - much can be accomplished before the full theory is in place. The mailing list nupic-theory would be the best place to discuss the algorithm stuff. --Subutai [1] <https://github.com/numenta/nupic.research/wiki/Overview-of-the-Temporal-Pooler> https://github.com/numenta/nupic.research/wiki/Overview-of-the-Temporal-Pooler On Wed, Jan 13, 2016 at 9:49 AM, cogmission (David Ray) < [email protected]> wrote: > Nice work Sebastian! > > I agree, Hierarchy as a concept is very important to neocortex emulation, > and to my knowledge Numenta has not released and is still working on an > implementation of this feature. The make-shift "hierarchies" that can be > composed using repeated combinations of the current algorithms are not > "true" hierarchies in the HTM sense (ascending regional constructs). This > being the case, for the time being repeated layers may not be as useful > (depending on your application of course), as they will be when the > hierarchy is actually part of the algorithm. > > Cheers, > David > > On Wed, Jan 13, 2016 at 11:35 AM, Sebastián Narváez <[email protected]> > wrote: > >> I had the same problem as you, and implemented a Layer class to play >> easly with different combinations of the Nupic modules (encoders, sps and >> tms). You can check it out here: >> >> https://github.com/larvasapiens/htm-teul/tree/master/Learning >> >> However, from my experiments, a one-level network >> (encoder->sp->tm->classifier) works better than a multi-level hierarchical >> approach with the current implementation of nupic. Or maybe I didn't used >> the correct parameters for my experiments. >> >> On Wed, Jan 13, 2016 at 11:43 AM, cogmission (David Ray) < >> [email protected]> wrote: >> >>> Hi Wakan, >>> >>> I while back a worked up a very raw example of connecting the algorithms >>> (in a very rough manner), in Python. Maybe this might be of some help? >>> (see attached) >>> >>> Cheers, >>> David >>> >>> On Wed, Jan 13, 2016 at 8:59 AM, 박진만 <[email protected]> wrote: >>> >>>> I did not use Network API. I made the structure by combining some >>>> example codes like hello_sp.py and hello_tm.py in >>>> https://github.com/numenta/nupic/tree/master/examples >>>> >>>> I made my midi encoder by combining 3 scalar encoders and 1 category >>>> encoder. I used FastCLAClassifier in nupic.bindings.algorithms >>>> >>>> I want to make an hierarchical structure of HTM, not using network API, >>>> but using those raw simple codes. >>>> >>>> Thank you. >>>> >>>> 2016-01-13 23:35 GMT+09:00 Wakan Tanka <[email protected]>: >>>> >>>>> Hello, >>>>> May I ask which framework are you using? OPF or Network API, from what >>>>> you've typed I guess this is Network API. I'm wondering how you made: >>>>> raw midi data -> Encoder -> Spatial Pooler -> Temporal Pooler -> CLA > >>>>> Classifier -> prediction >>>>> Also what encoder did you use? Have you followed some example codes? >>>>> >>>>> Thank you >>>>> >>>>> On 01/13/2016 02:47 PM, 박진만 wrote: >>>>> >>>>>> Hello, I'm newbie to NUPIC& NUPIC-mailing list. >>>>>> >>>>>> I'm working on training midi files(*.mid , a sort of music file) using >>>>>> low-level codes. >>>>>> >>>>>> Low-level codes mean just a raw sp and tp code, not network API or >>>>>> OPF. >>>>>> I prefer to use low-level codes because it's easier for me to modify >>>>>> the >>>>>> codes. >>>>>> I used a simple structure like this: >>>>>> raw midi data -> Encoder -> Spatial Pooler -> Temporal Pooler -> CLA >>>>>> Classifier -> prediction >>>>>> The result was quit awesome. the HTM successfully predicted the whole >>>>>> sequence with no error. >>>>>> >>>>>> Then I wanted to change the structure to be hierarchical like this : >>>>>> raw midi data -> Encoder -> SP1 -> TP1 -> SP2 -> TP2 -> CLA Classifier >>>>>> -> prediction >>>>>> >>>>>> but I cannot implement the structure because i don't know how the >>>>>> layer >>>>>> 1 and layer 2 is linked. I already watched >>>>>> "hierarchy_network_demo.py", >>>>>> but the code just tells us "UniformLink". >>>>>> What does the term "UniformLink" mean? >>>>>> >>>>>> I think it's gonna be a strange architecture if TP1's output(array of >>>>>> cells) becomes the input of SP2, because in this hierarchy, layer 2 >>>>>> (SP2 >>>>>> and TP2) will have bigger column dimension than those of layer 1, >>>>>> which >>>>>> is somehow weird. >>>>>> >>>>>> to sum up, my questions are : >>>>>> >>>>>> 1. How they are linked between layers. >>>>>> 2. Any hierarchy structure examples in low-level (not Network API, >>>>>> not OPF) >>>>>> >>>>>> Any comments would be very helpful. >>>>>> >>>>>> Thank you. >>>>>> >>>>>> >>>>> >>>>> >>>> >>> >>> >>> -- >>> *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 >>> >> >> > > > -- > *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 >
