SDR questions from HTM School videos

@rhyolight
Matt, thank you very much. Just found these videos, and they are very awesome!
I just finished episode 5 and I have some N00b theoretical questions:

You spend a lot of time talking about how SDRs get more resistant to errors and noise with larger and larger N, and they are being stored as sparse arrays, only storing the indices of the on bits; why not just assume an N of infinity? The chance of false positives pretty obviously approaches 0 as N goes to infinity. Union and Intersections could still be represented with a finite representation, though ‘Not’ would cause problems I suppose, since the inverse of “infinity choose finite” is an infinity…Calculus and Combinatorics…fun combo lol…

Second question, is there a similarity between SDR encoding and Bloom filters? They seem kind of similar; both are presented as probabilistic data structures, or at least used in a probabilistic way. The Union of SDR’s particularly seems to behave similarly to a bloom filter.

Thanks again!

1 Like

Thanks for visiting the forums! I appreciate your kind words. Please share the videos with friends and colleagues you think might be interested in HTM.

In many of the SDR dimensions we use in HTM (consider a typical spatial pooler input space of 600 choose 60), the capacity of the SDR can be considered infinity (2.77e+83 is astronomically large) . But if an SDR has a length of infinity, how could we compare that infinite structure to any other binary array? We would have to take a subset of the bits of each comparator array to get an overlap score relative to the sample size, and then what’s the point? We might as well have used a finite array in the first place.

Yes, as @floybix pointed out earlier.