I can't give a perfect number but Ian has a vision demo that might be a good starting point. Hopefully he can follow up with his numbers (I'm in a car driving to Oregon so can't look it up at the moment). You could use a small number of columns (like 10) if you are just having it recognize noisy versions of the digits. Something more complex might require hierarchy anyway.
And the cells per column controls the amount of temporal contexts the pattern can be learned in. You can probably put that at 10 without issue. Again, might be different for a complex problem. On Dec 26, 2013 11:17 AM, "Matthew Taylor" <[email protected]> wrote: > I have to say that I love the way you constructed the number cell data > set in code: > https://github.com/srgg6701/nupic/blob/gh-pages/www/cla/js/datasets.js > > I've never seen a data structure in code that actually resembles the > way the data would like in a UI. > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Wed, Dec 25, 2013 at 12:01 PM, Sergey Cleftsow <[email protected]> > wrote: > > Hi, everybody! > > I'm trying to use CLA in the simple example. Its prototype is available > > on-line (http://srgg6701.github.io/nupic/www/cla/numbers.html), but > there no > > any code for recognition yet, because now I haven't a starting assumption > > about which total number of columns and cells per single column would be > > more appropriate for this example. There is a some receptive area (2-D > > array) 16 х 20 elements. Inputs are "digits" stored in array within code. > > They may come in some modes like fixed positions, random offset and/or > > erosion. The speed of feeding and inputs limitation may be customized > also. > > > > I read in some papers that there were lots of experiments about this > > subject, but are there some mathematical (or another) basis to calculate > > necessary values. It will be great to hear advice! > > > > > > -- > > SY Sergey V. Cleftsow > > > > _______________________________________________ > > 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 >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
