You might want to look at the spatial pooler unit tests in NuPIC or HTM.Java:
- https://github.com/numenta/nupic/tree/master/tests/unit/nupic/research - https://github.com/numenta/htm.java/blob/master/src/test/java/org/numenta/nupic/algorithms/SpatialPoolerTest.java --------- Matt Taylor OS Community Flag-Bearer Numenta On Mon, Dec 14, 2015 at 10:40 AM, mraptor <[email protected]> wrote: > hi guys what are a good tests for testing Spatial pooler functionality ! > I'm implementing spatial pooler on my own for learning experience and want to > test if it works. > > I'm trying at the moment the Digits db from scikit learn. > What I'm doing currently is to train the system then predict SDR from > different sets of 0-9 digits > and compare their SDR's to see how many overlapping bits the different > SDR's have. > (i.e. overlapping_bits_count(set1-Zero, set5-Zero) ....) > I'm also interested how many of 2% of 2000 bits i.e. of 40 possible > bits how many are considered a good match. > > I haven't tried out but I would want to be able do some Clustering > and/or Classification !??! > It is just hard to wrap around my head how to do this with SDR as > compared to normal machine learning (X,y) pairs ! > Any ideas on that too ? > I have some but wanna hear of you can think of how do you do those > with SDR's rather than (X,y) pair. > F.e. every bit in SDR is a feature which you can group on (i.e. > cluster), you just have to figure out > based on the data which combination of bits occur most often and do > the clustering around them. > > One way to do Classification can be done if the "Spatial spooler" does > selective learning (by adding a "proximal"-input which allows learning > only at 1's to this input). > > and so on .... there seems to be myriad of ways of doing it. > > So any good test data and test scenarios...?? > > -------| http://ifni.co >
