Good point. I added a little blurb: "More precisely, is boosting causing the HTM to make use of all of its columns? Here’s an early glance at whether this is happening."
Yeah, it'd be interesting to see/do an in-depth study of boosting, compared to other algorithms (e.g. the one you mentioned: http://lists.numenta.org/pipermail/nupic-theory_lists.numenta.org/2015-December/003402.html ). On Sun, Mar 13, 2016 at 7:24 PM, Felix Andrews <[email protected]> wrote: > Hey, nice work. It's good to see that in detail. > > I think it's important to say, while you've showed boosting "working" in > the sense of getting neglected columns to become active, that should not be > confused with boosting "working" in the sense of leading to improved > prediction performance. Not sure if that's been tested by anyone (?) > > (f x) > > > On 14 March 2016 at 04:08, Marcus Lewis <[email protected]> wrote: > >> Hey all, >> >> I just posted: *HTM time series: Column overlaps and boosting* >> <http://mrcslws.com/blocks/2016/03/13/column-overlaps-and-boosting.html> >> >> This is a followup to my previous post >> <http://mrcslws.com/blocks/2016/02/15/htm-stacks-of-time-series.html>. I >> added some new d3 charts. The new charts answer questions that the previous >> charts caused me to ask. Also, the new charts cause me to ask a new set of >> questions. >> >> Feedback welcomed! Also, I'll be speaking informally at the Bay Area HTM >> Meetup <http://www.meetup.com/numenta/events/228433496/> tomorrow. It's >> a lightning talk on the topic: "See HTM from Multiple Angles". >> >> Marcus >> > > > > -- > Felix Andrews / 安福立 > http://www.neurofractal.org/felix/ >
