Hi Roy, In the CLA, "inhibition" is done at the column level, either globally or locally. The Spatial Pooler picks the 2% of columns with the highest activation potential from either the whole region (global) or sub-regions (local) and activates them. This mimics inhibition in the neocortex, in which high-potential (or most active) cells suppress the potential or activity of cells in neighbouring columns.
The overall effect is very similar, in that sparseness is imposed and the "best" columns are activated. If we can achieve very similar outcomes, then efficiency considerations justify the deviation from the actual process in the neocortex. A recent thread on this list discussed an alternate inhibition strategy for the CLA. This proposal involves the idea of high-potential columns imposing a "negative potential" effect on its neighbours. Initial tests suggest that this method is significantly faster than the current local "inhibition" algorithm, and it is also a closer analogue to the neocortex. Regards, Fergal Byrne On Fri, Sep 13, 2013 at 7:03 AM, Roy Gu <[email protected]> wrote: > Please check the video on the inhibition detail. > http://www.youtube.com/watch?v=LT3VKAr4roo > I suppose the video reports the truth. > > > 2013/9/13 Roy Gu <[email protected]> > >> Is inhibition a 'negative' weighted permanence? >> >> I know the HTM white paper said the permanence is 0.0 ~ 1.0 and the >> inhibition is to make sure the neuron activation is 'sparse locally'. >> However, why can't we consider the inhibition signal as a 'negative' weight >> of permanence? >> >> Could anyone help to provide some input? Do we have any information which >> supports me to make permanence as negative? >> >> -- >> Thanks, >> Roy >> > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > -- Fergal Byrne ExamSupport/StudyHub [email protected] http://www.examsupport.ie Dublin in Bits [email protected] http://www.inbits.com +353 83 4214179 Formerly of Adnet [email protected] http://www.adnet.ie
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
