With respect the Chadan's propposal on training 2 indeoendent regions... The regions would indeed learn to represent independently their respective streams of data, but how would you avoid the non-intentional overlaping when you feed the outputs from both regions into a higher level region? I think you would end up with lots of overlapping bits that have no semantical similarity.
On Mon, Mar 28, 2016 at 1:54 PM, Chandan Maruthi <[email protected]> wrote: > Phil, > I think little work as been done on hierarchy at present. I personally am > interested but haven found time to work on hierarchy. > > > Chandan > > On Mon, Mar 28, 2016 at 9:38 AM, Phil Goddard <[email protected]> wrote: > >> Chandan, >> >> >> thanks for your thoughts. >> >> >> I am assuming that I'll need to use some sort of hierarchical >> (multi-layer) model, and have not found much info on that approach. >> >> Was hoping that someone might be able to point me towards some papers or >> examples. >> >> >> Phil. >> >> >> ------------------------------ >> *From:* nupic <[email protected]> on behalf of Chandan >> Maruthi <[email protected]> >> *Sent:* Monday, 28 March 2016 3:43 PM >> >> *To:* [email protected] >> *Subject:* Re: Multi-input and multi-level models >> >> Phil, >> I can think of 2 approaches >> 1.) Normalize both input data streams to the same time series [may be >> with averaging out the changes during the difference in time stamps] and >> then feed it as a 1 time series. >> 2.) Use a hierarchical model to train 2 independent regions with the >> separate time series data which feeds into 1 higher level region. This >> should allow for the formation of independent low level representations of >> the data. And they may connect at a higher level region. [This will be more >> experimental and not much work has been done with hierarchy] >> >> Chandan >> >> On Mon, Mar 28, 2016 at 8:13 AM, Matthew Taylor <[email protected]> wrote: >> >>> Phil, I have an example project that will converge multiple River View >>> data streams into one model properly with NuPIC. It is called Menorah, >>> check it out here: https://github.com/nupic-community/menorah >>> >>> If nothing else, it is an example of converging disparate data sources >>> into a multi-field model. >>> >>> --------- >>> Matt Taylor >>> OS Community Flag-Bearer >>> Numenta >>> >>> On Mon, Mar 28, 2016 at 7:34 AM, Phil Goddard <[email protected]> >>> wrote: >>> >>>> Hi Marcus, >>>> >>>> >>>> thanks for the links, I'll take a look at them shortly. >>>> >>>> >>>> In terms of the hotgym example, I think of that as really being 1 time >>>> series input rather than multiple inputs. >>>> >>>> >>>> One thing I need is to be able to input multiple time series'. >>>> >>>> In particular, inputs that arrive at different time intervals. >>>> >>>> (The inputs can't be handled by aligning them with just one time vector >>>> - each input needs it's own time vector.) >>>> >>>> >>>> Thanks >>>> >>>> Phil. >>>> >>>> ------------------------------ >>>> *From:* nupic <[email protected]> on behalf of Marcus >>>> Lewis <[email protected]> >>>> *Sent:* Monday, 28 March 2016 5:49 AM >>>> *To:* [email protected] >>>> *Subject:* Re: Multi-input and multi-level models >>>> >>>> Hi Phil, >>>> >>>> Multiple inputs are pretty easy with the OPF (i.e. the CLAModel). For >>>> example, hotgym >>>> <https://github.com/numenta/nupic/blob/master/examples/opf/clients/hotgym/simple/model_params.py#L66> >>>> uses the timestamp and the current power consumption. >>>> >>>> For multi-level models, you'll probably want to use the Network API. >>>> Subutai gave a recent talk: https://www.youtube.com/watch?v=g9yS9zFt3dM . >>>> Here's a demo that uses multiple levels: >>>> https://github.com/numenta/nupic/blob/master/examples/network/hierarchy_network_demo.py >>>> >>>> Strictly speaking, from the Network API's perspective, the CLAModel >>>> only has one input, since it concatenates >>>> <https://github.com/numenta/nupic/blob/master/src/nupic/frameworks/opf/clamodel.py#L1108> >>>> the >>>> encodings via a MultiEncoder, but that's just an implementation detail. I >>>> recently created a demo that puts multiple inputs into a Network: >>>> https://github.com/numenta/nupic/blob/master/examples/network/core_encoders_demo.py >>>> >>>> >>>> Hope that helps! >>>> Marcus >>>> >>>> On Sun, Mar 27, 2016 at 10:06 PM, Phil Goddard <[email protected]> >>>> wrote: >>>> >>>>> >>>>> I'm looking for an example (or examples) of using NuPIC with either >>>>> (or both of) multiple inputs and multiple levels. >>>>> >>>>> >>>>> With multiple inputs, the nearest example I can find is the NY Taxi >>>>> example. >>>>> >>>>> However the technical paper I have indicates that the 3 inputs are >>>>> aggregated into one input (via competitive polling) before being fed into >>>>> the model. >>>>> >>>>> Can anyone tell me if it possible to have multiple inputs? >>>>> >>>>> Or do multiple inputs have to be processed into one input as per that >>>>> example? >>>>> >>>>> >>>>> I can't find the code for the NY taxi example (in the NuPIC GitHub >>>>> repository). >>>>> >>>>> Is it available, and if so where? >>>>> >>>>> >>>>> Also, is it possible to develop multi-level models? >>>>> >>>>> If anyone can point me at any technical description of such models, or >>>>> a code example, I'd appreciate it. >>>>> >>>>> >>>>> thanks >>>>> >>>>> Phil. >>>>> >>>>> >>>>> >>>>> >>>>> >>>> >>> >> >> >> -- >> Regards >> Chandan Maruthi >> >> > > > -- > Regards > Chandan Maruthi > >
