I have few questions. 1) I have good understanding on how to encode scalar values. I’m working on the vision problem at the moment and stumbled upon the coding of the video signal. The size of the sensor is relatively small (8x8 pixels), however I would like to ensure that the the value of each pixel is part of the coding process (0-255). Does it mean that in order to achieve sparsity I should treat each pixel as a scalar encoder, thus resulting in 64 encoders?
2) Given the idea of sparse distributed representations, copied from the human brain, how is vision signal coded and becomes sparse in the V1? Does approximately 1 million afferent inputs from the retina get sparse? Do this inputs get coded first to represent features (gabor filters, direction of the feature’s motion etc) and then become sparse? Is the “spatial pooler” of V1 just so much bigger then the afferent signal size (1 million afferent signals vs 50 million mini columns in V1) ? 3) I’ve read that both motor signal and sensory signal are used in order to achieve sensory-motor transition representation. If this idea is overlaid on the HTM framework, how do these two inputs are added/mixed in a single region? This can be a question of multiple sources feeding the same region as well, in general. How Layer 4 knows the difference between sensory and motor signals to learn sensory-motor transitions? Sincerely, Dennis S. Sedov > On Dec 8, 2014, at 10:49 AM, Michael Klachko <[email protected]> wrote: > > My question: is there a plan to update HTM white paper on Numenta website? > It seems like it hasn't been touched since 2011. > > On a related note, any plans to write a second edition of "On Intelligence"? > This is still the main source of information people refer to when they > discuss HTM, and it's missing important ideas developed in the last decade > (like SDR). > > > > > > > On Mon, Dec 8, 2014 at 10:34 AM, Matthew Taylor <[email protected] > <mailto:[email protected]>> wrote: > Any questions you leave here, I'll have prepared in a Wiki page for the > office hour for reference. We'll take live questions first, then Q&A > questions, then address the questions on the wiki. > > Thanks, > > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > On Mon, Dec 8, 2014 at 10:09 AM, Chandan Maruthi <[email protected] > <mailto:[email protected]>> wrote: > Btw that was a question for the office hour :-) > > On Mon, Dec 8, 2014 at 10:08 AM, Chandan Maruthi <[email protected] > <mailto:[email protected]>> wrote: > I have a question. > > I see we have a few frameworks within numenta. Network API, OPF etc. > Whats the general direction going forward. Where you would suggest one > framework vs another. > Is there more support for specific frameworks in the product roadmap . > > > Regards > Chandan Maruthi > > > On Mon, Dec 8, 2014 at 7:43 AM, Matthew Taylor <[email protected] > <mailto:[email protected]>> wrote: > Hello NuPIC, > > We'll be holding another office hour today. > > https://plus.google.com/b/100642636108337517466/events/c89fj875svdde9si3dffkq373s0 > > <https://plus.google.com/b/100642636108337517466/events/c89fj875svdde9si3dffkq373s0> > > We don't have a topic set yet, but I'm open to suggestions. > > Please contact me if you want to join directly and ask a question. Otherwise, > you can use the Hangout Q&A interface to ask questions or respond to this > email with your question or comment. > > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > >
