Hi Traun, Your basic approach seems quite reasonable as a first step. I would definitely look at Kevin's tutorial in the link that Matt sent out as it might answer many of your questions about the class. (And don't use the oldpy implementation. :-)
For doing classification, take a look at: py/nupic/algorithms/KNNClassifier.py There are some examples of using it here: tests/integration/py2/nupic/algorithms/knn_classifier_test/ It is lower level than the CLAClassifierRegion. For distance metrics, with SDRs we usually use an overlap distance metric (count the number of bits that are common) rather than Euclidean distance. If you are just using a spatial pooler it shouldn't matter too much, but it does matter a lot once you start using the temporal pooler. I think your assumptions about close matches are reasonable. You might not get an exact match unless the inputs are really really close. In particular, the spatial pooler won't do too well at learning general invariances (and it's not supposed to). For example, if you shift the image by one or two pixels you might get a very different output SDR. I expect you will have to understand and tune some of the parameters. I think you'll learn a lot in the process. Look forward to hearing more as you work on this! --Subutai On Tue, Mar 4, 2014 at 11:05 AM, Matthew Taylor <[email protected]> wrote: > Take a look at the links in this wiki we just created: > https://github.com/numenta/nupic/wiki/Introduction-to-the-Algorithms > > Kevin is working on a tutorial that might help you use the SP. > > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Mon, Mar 3, 2014 at 6:15 PM, Traun Leyden <[email protected]>wrote: > >> >> I'm trying to figure out how to write a digit recognizer implementation >> using a spatial pooler and had a few questions. >> >> The data would be similar to mnist, but I would start with a much simpler >> data set. The input vector would be a 1d vector of 784 input elements that >> represented the 2d image array of 28x28 pixels. As a simplification, the >> elements of the vectors would be 0 or 1 (as opposed to greyscale values >> used in mnist). >> >> The data sets I was planning to use: >> >> * For training (online learning), I would create ideal versions of the >> 0-9 digits >> >> * For testing (eg, online learning turned off), I would create noisy >> versions of those same digits >> >> >> The approach I was planning to take: >> >> 1) Create a spatial pooler instance >> >> 2) Turn on online learning >> >> 3) Repeatedly present ideal input vectors of 0-9 to the spatial pooler >> >> 4) Record the final SDR's of each input vector >> >> 5) Turn off online learning >> >> 6) Present noisy input vectors of 0-9 to the spatial pooler >> >> 7) Find the "closest" SDR recorded in step 4, and consider that the >> inferred SDR of the spatial pooler. I would compare the inferred SDR with >> a known expected SDR, which would be used to calculate the overall error. >> >> >> Does that sound like a reasonable approach? If not, what's a recommended >> approach to go about building a simple digit recognizer using the spatial >> pooler? >> >> Also I had some specific questions about how to use a spatial pooler: >> >> * Are there any examples that directly use a spatial pooler that I can >> look at? The closest I could find was >> https://github.com/allanino/nupic-classifier-mnist.git, but that uses >> OPF and I want to go straight to the lower level code and use the spatial >> pooler. >> >> * After looking at the code for the spatial pooler implementations, the >> "py" and "cpp" implementations don't seem to return any values. Only the >> "oldpy" implementation seemed to return a list of the active columns. >> Here's the method signature of the cpp spatial pooler I'm looking at in >> spatial_pooler.cpp: void SpatialPooler::compute(UInt inputArray[], bool >> learn, UInt activeArray[]) { } If they don't return any values, how are >> they supposed to be used? >> >> * Is there anything special to do with regards to recognizing 2d image >> patterns? Or will the columns naturally form 2d receptive fields of the >> input vector? >> >> * Since I want to basically classify digits, should I be looking at the >> CLAClassifierRegion for guidance? The idea I came up with for comparing >> the SDR's feels a bit janky, and it seems like there should be a cleaner >> way. Eg, something helps take care of the classification work but that's >> still lower level than the OPF? >> >> * I am expecting to see exact matches for SDR's when presenting input >> vectors that only have a little bit of noise, and "close" matches for input >> vectors that have a lot of noise. Is this a correct assumption? Is there >> any built-in machinery for measuring how close two SDR's are to each other? >> >> >> Thanks in advance for any help, I'm really excited to get started >> building this. I'm planning to contribute it back in a pull request once >> it's working. >> >> >> >> _______________________________________________ >> 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
