Hello, As Marek said, you would not be taking advantage of the temporal aspects of HTM. However, if you want something similar to HTM which can be used for image recognition, look into sparse coding (aka SDR) literature. Stuff like K-sparse autoencoders function similar to HTM's spatial pooler, and are well suited to image recognition (indeed, K-sparse autoencoders get very good performance on MNIST and other datasets). You will need to put a classifier on top of the sparse coder, but this can often just be a linear combination and it will work well.
Link to K-sparse autoencoder paper: http://arxiv.org/abs/1312.5663 Regards, Eric Laukien On Wed, Feb 24, 2016 at 8:07 PM, Marek Otahal <[email protected]> wrote: > You can use HTM for images, you just won't take any advantage from the > temporal pooling (TP), but you can put the properties of SDRs and spatial > pooler SP to a good use. > There are 2 more recent papers on HTM and vision, one by a Slovak > researcher, and one by Eric(k)? Both utilizing SP. > > I have some interest in the nupic.vision :) > > On Wed, Feb 24, 2016 at 8:51 PM, Sergey Alexashenko < > [email protected]> wrote: > >> That's part of it. Frankly, I am not entirely sure that NuPIC is fast >> enough to work well with images in its current implementation (although I'd >> love to hear educated opinions on this one). 2048 columns doesn't seem >> enough and scaling is exponential. >> >> Has anyone here actually worked with nupic.vision? ( >> https://github.com/numenta/nupic.vision) >> >> On Wed, Feb 24, 2016 at 12:34 PM, cogmission (David Ray) < >> [email protected]> wrote: >> >>> From what I understand, we have to wait for "Hierarchy" to be added back >>> into the algorithm before significant work can be done with HTMs for vision? >>> >>> On Tue, Feb 23, 2016 at 9:39 PM, Sergey Alexashenko < >>> [email protected]> wrote: >>> >>>> If you look at the citations of the paper you just linked, it cites >>>> this >>>> <http://www-edlab.cs.umass.edu/cs691jj/hawkins-and-george-2006.pdf> 2006 >>>> report. Back then, Numenta used a very different algorithm from what it >>>> uses now, even though the name remains the same. I don't know how well the >>>> old algorithm worked in image recognition, just be aware that it is not the >>>> same as NuPIC, the current implementation of HTM. >>>> >>>> On Tue, Feb 23, 2016 at 10:26 PM, Николай Климов <[email protected]> >>>> wrote: >>>> >>>>> Hello guys. I'm new to HTM and I have a question about object >>>>> recognition in images. Every source I've read about HTM said it's a bad >>>>> idea because it's not a temporary data. But I've just read master thesis >>>>> by >>>>> Vincenzo Lomonaco ( >>>>> http://amslaurea.unibo.it/9095/1/Vincenzo_Lomonaco_tesi.pdf ) where >>>>> in some tests HTM beats CNN. Why it works? I ask this because I want to >>>>> try >>>>> implement face recognition task on HTM and wondering is it a good idea? >>>>> >>>>> >>>> >>> >>> >>> -- >>> *With kind regards,* >>> >>> David Ray >>> Java Solutions Architect >>> >>> *Cortical.io <http://cortical.io/>* >>> Sponsor of: HTM.java <https://github.com/numenta/htm.java> >>> >>> [email protected] >>> http://cortical.io >>> >> >> > > > -- > Marek Otahal :o) >
