Hi Mark, > On Wed, Aug 7, 2013 at 5:22 PM, Francisco Webber <[email protected]> wrote: > Still concerning the pass-through encoder: > At various occasions I have noticed that the "pass through" functionality > corresponds to skipping the spacial pooler. > In my current understanding this would not be correct. > It is true that the output of the spatial pooler is in SDR format. But it > cannot be implied from there that the the spacial pooler actually SDRizes the > input. > If an SDR is feed in the SP that (randomly) subsamples it, the resulting > output will still be a valid SDR (through the noise resistance property of > the initial SDR). > If a non SDR bit pattern is subsampled the result will be looking like an SDR > while having lost the semantic context of the input data (lacking the NR > property. > Yes, I believe a SP "just creates" a SDR, like Fergal said, this happens > "automatically" in neurons through inhibition. You then agree with me that this "creation" can just be done with an SDR input? > > Another problem that I have on the conceptual level, when assuming that SDR > data doesn't need the SP stage is that in a hierarchy of layers the output > data of every layer would have the SDR format. If it would be fed into the > next higher level, one could skip the SP in all subsequent higher levels , > leaving the question why there is an SP at all. > Here is the second role of a SP, it subsamples/compresses the output. So for > a hiearchy of regions, say 10 regions have output of 1000 bits each, now that > makes 10000 for the new input to the higher region. SP can shrink those 10k > bits to say 1000 and feed it the the next TP. Yes this seems clear to me, this is one main purpose of hierarchies to "integrate" the data flow. > > As SP is trained, it also handles noice at the input.
THe noise-resitance I mentioned referred to the (desirable) properties of SDRs to allow missing or wrong data while still maintaining the semantics of the initial pattern. Thank you for helping to understand Francisco
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
