You are correct that a separate scalar encoder for each pixel would not be a good method. You would want to create a vision-specific encoder which may do a number of preprocessing techniques to capture the semantics of the data. You would also want to use topology in the SP (currently the default is global inhibition). And I believe that hierarchy is somewhat important to vision problems, although may not be necessary for some problems (I don't really know).
This is a pretty big task in general. We have been talking about putting up details around some domains like vision and NLP to help guide research so we will follow up if we can get something like that together for vision. On Tue, Aug 13, 2013 at 8:21 AM, Hideaki Suzuki <[email protected]> wrote: > Hello people, > > Today, I've got one idea in my mind. Do you have any opinion or interest > for this? > > Nupic uses scalar encoder to feed numbers to SP, IIUC. > > If we want to feed an image to SP, each pixel has an integer value and > the number of pixels are large. It sounds not so easy to me that we > convert > million pixel values through scalar encoder and feed them to SP. > > Can dithering work as the encoder? I mean this. > http://en.wikipedia.org/wiki/Dither > > Our brain recognizes a black & white binary image as if it is a grayscale > image > after dithering. Does our retina convey the intensity of light in a > similar way of > dithering...? > > # I was a bit away from ML due to a business trip. I hope I can be more > active now. :) > > Best Regards, > Hideaki Suzuki. > > _______________________________________________ > 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
