Hi Ozgur, Face detection is definitely a more difficult problem than the optical character recognition one that I'm working on. If you want it to distinguish between different people I think that would be very challenging. If you just want it to tell you if there is any face in a picture that could be a good project and is similar to what Vitamin D was doing with Numenta in the early days. Vitamin D was purchased by Sighthound, http://www.sighthound.com.
As far as being able to recognize patterns at different locations in an image, it is true that the spatial pooler is not good at this by itself, but I don't see this as a problem with the spatial pooler or even as a particularly tough problem. When I look at a picture I can only recognize objects in the image that are in the center of my field of view. If I want to identify an object in the upper right corner then I have to look at the upper right corner thereby changing my field of view. So I view this problem as one of how you process the image and present it to the spatial pooler rather than as a problem with the spatial pooler. On Tue, Jul 8, 2014 at 6:28 AM, R. Özgür Aksu <[email protected]> wrote: > Hey Jim, > > When Numenta first came out with their product, I mean when it was called > HTM, they tried some visual pattern recognition. I think Jeff had said that > they saw various degrees of success, but I don't know the details. When I > saw that I remember wanting to try it out for object recognition but > especially face detection. However, the current setup of feeding in each > pixel to the spatial pooler makes me think that it will not be enough. I > don't know this for a fact, but it was enough to discourage me from even > trying. The thing is as I understand it, the spatial pooler is great at > fuzziness, but it is not that great at seeing a pattern at, say, location > (10, 10), and recognizing a similar pattern later at location (50, 10). > Correct me if I'm wrong but that is what I assumed and so didn't get into > it. > > Ozgur > > > On Tue, Jul 8, 2014 at 12:24 AM, Jim Bridgewater <[email protected]> wrote: >> >> Matt, >> >> On a separate subject I've been thinking it would be super cool to >> have some way to post a snap shot of all the code and data required >> for each set of results so if someone is interested after looking at >> the results they could run the code themselves. I'm not sure what the >> best way to do this is. Git tags or a webpage or something else. Any >> ideas? >> >> On Mon, Jul 7, 2014 at 2:20 PM, Jim Bridgewater <[email protected]> >> wrote: >> > Great! I'm glad you gave it a try Matt. >> > >> > I ran into that JPEG decoder issue on the Amazon EC2 image that Ian >> > put together. It looks like PIL and pillow depend on libjpeg-dev for >> > this so it has to be installed when PIL / pillow is compiled. >> > >> > >> > http://stackoverflow.com/questions/8915296/python-image-library-fails-with-message-decoder-jpeg-not-available-pil >> > >> > >> > >> > >> > >> > On Mon, Jul 7, 2014 at 8:41 AM, Matthew Taylor <[email protected]> wrote: >> >> Thanks Jim. I was able to run your demo, but I had to recompile PIL >> >> because >> >> the "jpeg decoder" was missing. >> >> >> >> --------- >> >> Matt Taylor >> >> OS Community Flag-Bearer >> >> Numenta >> >> >> >> >> >> On Sun, Jul 6, 2014 at 11:09 PM, Jim Bridgewater <[email protected]> >> >> wrote: >> >>> >> >>> Here's the version of this experiment using an SP with 1024 columns >> >>> and 16 active columns on a data set that contains 16 characters (0-9, >> >>> A-F). >> >>> >> >>> The results are similar in that setting increment for active columns, >> >>> synPermActiveInc, equal to the connection threshold, synPermConnected, >> >>> tends to produce the best results. However, one exception was that >> >>> setting both increment and decrement values to 0.1*synPermConnected >> >>> produced 100% accuracies for all but the highest threshold values. >> >>> This was not the case for the SP with only one active column. >> >>> >> >>> This time I included pictures of the permanences for the best and >> >>> worst configurations for comparison. For the worst configuration many >> >>> columns have permanences far above zero for inputs that are not >> >>> active for any of the input images. Maybe these columns were never >> >>> active and therefore those permanences were never decremented? I >> >>> guess that is another experiment to run...recording active columns to >> >>> see how many get used. >> >>> >> >>> On Sat, Jul 5, 2014 at 8:23 AM, Ian Danforth <[email protected]> >> >>> wrote: >> >>> > >> >>> > >> >>> > >> >>> > On Fri, Jul 4, 2014 at 9:31 PM, Jim Bridgewater <[email protected]> >> >>> > wrote: >> >>> >> >> >>> >> Hi Ian, >> >>> >> >> >>> >> You are correct. Thanks for the feedback, I've updated the >> >>> >> introduction and attached the revised document. >> >>> >> >> >>> >> There is still the question of whether these results will hold with >> >>> >> more active columns and I plan to run that experiment and post the >> >>> >> results soon. I am planning to use your Amazon EC2 image for this >> >>> >> since these jobs can take some time for large numbers of columns. >> >>> >> >> >>> > >> >>> > Awesome! Let me know if you end up installing tools that you think >> >>> > should be >> >>> > part of the default AMI. >> >>> > >> >>> >> >> >>> >> For pooling, I want to duplicate the characters in the data set >> >>> >> using >> >>> >> different fonts and see if I can get the spatial pooler to >> >>> >> generalize >> >>> >> between the different fonts. >> >>> >> >> >>> >> >> >>> > >> >>> > Looking forward to those results! >> >>> > >> >>> > Ian >> >>> > >> >>> > >> >>> > >> >>> >> >> >>> >> >> >>> >> On Fri, Jul 4, 2014 at 6:43 PM, Ian Danforth >> >>> >> <[email protected]> >> >>> >> wrote: >> >>> >> > Jim, >> >>> >> > >> >>> >> > To be clear you have 186 training examples and 256 columns, 100% >> >>> >> > potential >> >>> >> > pool and you have 1 column on at a time, so you would expect 100% >> >>> >> > accuracy. >> >>> >> > Correct? You might want to note this up front. Something like >> >>> >> > "this >> >>> >> > should >> >>> >> > be an easy task for the spatial pooler, however certain parameter >> >>> >> > configurations were found to be problematic." >> >>> >> > >> >>> >> > Also I know you know this but for others reading along, to do >> >>> >> > "pooling" >> >>> >> > the >> >>> >> > number of training examples needs to be at least > than number of >> >>> >> > columns >> >>> >> > and preferably >>. Otherwise each column will perfectly fit a >> >>> >> > single >> >>> >> > example. Also you need to test on a separate set of data to see >> >>> >> > how >> >>> >> > well >> >>> >> > the >> >>> >> > model generalizes. >> >>> >> > >> >>> >> > Ian >> >>> >> > >> >>> >> > >> >>> >> > On Fri, Jul 4, 2014 at 12:29 AM, Jim Bridgewater >> >>> >> > <[email protected]> >> >>> >> > wrote: >> >>> >> >> >> >>> >> >> Hi everyone, >> >>> >> >> >> >>> >> >> I'm working on a NuPIC vision toolkit for the Season of NuPIC >> >>> >> >> this >> >>> >> >> summer and I've created a GitHub repo for it called >> >>> >> >> nupic.vision. >> >>> >> >> The >> >>> >> >> URL is: >> >>> >> >> >> >>> >> >> https://github.com/baroobob/nupic.vision >> >>> >> >> >> >>> >> >> There is a demo for those who want to try it out. >> >>> >> >> >> >>> >> >> I've attached the write up of results from using this toolkit to >> >>> >> >> investigate the effects of the synapse connection threshold, >> >>> >> >> synPermConnected, the permanence increment for active synapses, >> >>> >> >> synPermActiveInc, and the permanence decrement for inactive >> >>> >> >> synapses, >> >>> >> >> synPermInactiveDec, on image recognition accuracy and the amount >> >>> >> >> of >> >>> >> >> training required. >> >>> >> >> >> >>> >> >> -- >> >>> >> >> Jim Bridgewater, PhD >> >>> >> >> Arizona State University >> >>> >> >> 480-227-9592 >> >>> >> >> >> >>> >> >> _______________________________________________ >> >>> >> >> 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 >> >>> >> > >> >>> >> >> >>> >> >> >>> >> >> >>> >> -- >> >>> >> Jim Bridgewater, PhD >> >>> >> Arizona State University >> >>> >> 480-227-9592 >> >>> >> >> >>> >> _______________________________________________ >> >>> >> 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 >> >>> > >> >>> >> >>> >> >>> >> >>> -- >> >>> Jim Bridgewater, PhD >> >>> Arizona State University >> >>> 480-227-9592 >> >>> >> >>> _______________________________________________ >> >>> 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 >> >> >> > >> > >> > >> > -- >> > Jim Bridgewater, PhD >> > Arizona State University >> > 480-227-9592 >> >> >> >> -- >> Jim Bridgewater, PhD >> Arizona State University >> 480-227-9592 >> >> _______________________________________________ >> 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 > -- Jim Bridgewater, PhD Arizona State University 480-227-9592 _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
