I have built many Encoders that encode/decode single SDR f.e. sdr = encoder.encode(55) encoder.decode(sdr) It is easy to write encoder to do this : union = encoder.encode([11,22,55]) the problem is : encoder.decode(union) You can do it by enumeration, but is not feasible ! Ima…

I can think of: Nearest Neighbor indexing/search Associative memory with SDRs Each has its own costs, but consider yet a more serious issue - assuming problem space become more complex and you use N encoders for different modalities/input streams, and overlaps may contain SDRs encoded by differen…

u can use something like ths [image] External Similarity Engineering As I mentioned many times in this forum Similarity using OVERLAP or for that matter hamming distance, dot product, cosine similarity OR any other operation on any type of Vectors (SDR,Kanerva hyper vecto…

For what? Orthogonality has its own undeniable virtues, but Similarity is good when you want an approximate value in case you fail to decode an exact match Another problem I have with a naive encoder that just overlaps sdrs for several values is that there is lost any semantic about what these v…

Either way, I also think the idea to use “SDR triplets” to encode/describe arbitrary relationships between arbitrary pairs-of-things might be very fruitful. Maybe. I have yet to figure out how are SDRs more appropriate than explicit graph description, other than a warm “sdr is good” feeling

[image] mraptor: how will you develop encoder for unions ? I guess, do you have to? Or for a given column mapping’s potential pool, let it reference those multiple encoding spaces instead of trying to union them. So: encodingA encodingB encodingC and column 1’s potential connections would…

This goes to the larger question of feature binding. I like introducing triplets of (feature relation feature) at multiple levels of representation. Of course, this requires introducing layers of representation in your model and the whole infrastructure that goes with this. I will add that DL dep…

I’m asking with reference with my prev. post … where a TM outputs a union of SDRs and Encoder feeds TM directly w/o SP

[image] cezar_t: Except for the case you use different numbers to represent unique IDs, eg the union SDR represents three identities (words, names, pointers) and not three smoothly changing scalar values. Then you prefer 55 to be orthogonal to 54 yes thats what i meant… merging different dims…

By the way, simple enumeration works very well! If you ignore the performance, that is. The SDR of 11: [image] The SDR of 22: [image] The SDR of 55: [image] The union: [image] Overlap counts between the all 1000 SDRs and the union(bottom: zoomed to only show up to 99): [image] The indice…

So this is what I’ve come up with: You first measure overlaps between the union and clusters of encodings. A cluster is made up of the union of randomly chosen encoded SDRs. You can think of a cluster as a set of dendritic segments. It encodes separate independent SDRs. (Even a temporal memory c…

Encoding/Decoding union of SDRs

Engineering

mraptor January 26, 2022, 4:12pm 3

u can use something like ths

1 Like

Topic		Replies	Views
I've found a way to decode SDRs Engineering	18	2169	July 22, 2017
Questions regarding SDR encoding Engineering	9	3451	September 14, 2016
How get SDR of word Engineering	37	1494	May 17, 2018
Implementing SDR long term memory? Implementations memory	9	1669	March 6, 2022
Most frequent SDR overpowering all other bits Engineering encoders , question , community	21	713	April 15, 2024

Encoding/Decoding union of SDRs

Related topics