Any idea if there is systematic way to transform/project/permute SDRs in such a way so that you “reverse the overlap”, so that similar SDR become dissimilar and vs versa … remember it has to be systematic way
I would guess some form of encryption applied to SDRs should do it.
A good encryption scheme transforms two slightly different blocks into two very different blocks, and allows decrypting back too.
But you have to give up noise robustness since it is totally opposite to the goal of exacerbating differences.
I guess this is not what you (we) want. We want to have some good robustness to random bit flips.
And I think I understand why. Sometimes when we have lots of similar patterns we want to “enhance” differences in order to make more visible in what ways they differ.
I guess some form of “SDR compression” could be possible, but my feeling is it will have to be learned against actual real world data. I don’t think it can be generic or dataset agnostic
implementing TM again …when bursting the new learned pattern differ very slightly… so the memory will fill in clusters… cause data SDRs are not uniformly distributed
Tested with random uniformly generated SDRs… by extrapolating a grid of size 10000x10000 bits ~ 11MB should be able to store upto 1 000 000 transitions. (the test show 100 fold less than the total 100 mln bits of Ts, where I will have max upto FIVE false positives)
So if I can decluster I hope to get closer to this ideal capacity
I have been thinking of SDRs as like GUIDs: unique, distinctive, but with noise resistance built in. The only way to adapt an SDR output to some input is by triggering an intermediate column to do the conversion. Mathematical transformations on SDRs of whatever kind don’t seem to sit right with that model.