The separation of SP and TP is an implementation detail but it makes it much easier to reason about the code. And I wouldn't say that the SP simply subsamples its input. The white paper may be a good resource for exactly how it works but it is possible to have a larger SDR output from the SP than it receives as input. For instance, if you choose 2000 columns with 40 active for the CLA and you have a data set with a timestamp and a scalar predicted value, it is likely that the SDR from the encoders will be much less than 2000 bits but the SP will still output a valid 2000 bit SDR.
And there is a little more going on in choosing which columns are active (we call them coincidences in the SP code rather than columns, btw). Boosting, for instance, makes columns that haven't been active for a while more likely to become active in order to ensure that all columns learn to represent something. On Wed, Aug 7, 2013 at 8:53 AM, Francisco Webber <[email protected]> wrote: > Thanks for the explanation Fergal, > > The Spatial Pooler is an artefact of NuPIC; > > Does this mean that it is not used anymore in the current state? > > it simulates the way a layer of columns becomes activated by feedforward > inputs. > > But this assumes then that the FF-inputs receive "true" SDR data, right? > > In essence, the algorithm is: > > Get all columns to check their feedforward dendrite and generate an > activation score. > Pick the top 2% of columns based on activation score and activate them. > Update the permanences on the feedforward dendrites. > > > This makes sense to me (assuming the input is SDR) > > In the neocortex, the columns do this themselves, and they pick the top 2% > (may vary) by higher-score columns inhibiting nearby lower-score columns. > They also update their own synapses themselves. > > So the SP is simulating the inter-region communication as well as the > single-region sensory communication. > > > Thank you for clarifying > > Regards > > Francisco > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
