Preferred way to pass recognized sequences from Temporal Memory to a "higher HTM layer"?

Engineering NuPIC
,
2

Hi,

IMO hashing is a bad idea. It is certainly not what the brain does. A problem with hashing an SDR is that the hash depends on the exact value of the SDR, so if even a single cell’s activity is different then the hash will be totally different. Hashes don’t account for the semantic similarities between their inputs.

This article describes a good solution to your problem.
(PDF) Learning Invariance From Transformation Sequences. Foldiak, 1991

I’ve personally tested this solution and you can see my results at the end of the video I recorded about it. In the video: I created a temporal memory (TM) and fed words into it, one letter at a time. Then I used this method to make an SDR that represents the current word, given only the output of the TM.

1 Like