How to encode large vector as input to SP and HTM

[aamir-gmail]

I have a large feature vector (128, 300, 600 and 1200 D), is there a way to encode them using RSDE
for input to SP.

[breznak]

(128, 300, 600 and 1200 D)

do you mean n-dimensional vector, or 1D with that many elements?
If 1D, and each element means a feature. Make a SDR for each element (and it's RDSE encoder). Then concatenate() the SDRs into one, feed it to SP.

[aamir-gmail]

I have 128 elements (1D with many elements) vector as in [x1,x2,x3,............], currently, I am using TSNE dimensional reduction to 3 1D elements. as with many dimensional reductions techniques I am sure some of the original properties are lost in this reduction. I want to get your opinion on the best way forward.

[breznak]

I want to get your opinion on the best way forward.

You'll need to test this for your specific task. In general. HTM's SpatialPooler should perform well in that matter.
Please let us know your results, it'll be interesting.

[aamir-gmail]

The way I have gone about doing this is available in my GitHub repository FYI https://github.com/aamir-gmail/htm-vision-encoder I have used TSNE to reduce the dimensions of 128 D vector to 3 and 4th dimension is the norm of the 128 D vector, so we have 4 inputs to SP encoded VIA RDSE Cheers Aamir

…

On Sat, Apr 11, 2020 at 1:05 AM breznak ***@***.***> wrote: Closed #782 <#782>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#782 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJA2ECPSVDPRWQMNLE3BTUDRL4YS7ANCNFSM4LIFROQQ> .

-- Kind Regards Aamir Mirza

[breznak]

thanks for the heads up, Aamir!

I have used TSNE to reduce the dimensions of 128 D vector to 3

do you have a compirasion of tSNE (or other dimensionality reduction methods, PCA) to HTM's spatial-pooler?

Btw, for images, you could try a biological RetinaEncoder (as a working PR in this repo)

[aamir-gmail]

Yes, I tried PCA as well as UMAP, however, TSNE worded best, please clone the GitHub repo see for your self. However, the problem I faced with TSNE is that it is slow and there is very much an upper limit on TNSE on how much data can be encoded.

…

On Tue, Apr 14, 2020 at 6:52 PM breznak ***@***.***> wrote: thanks for the heads up, Aamir! I have used TSNE to reduce the dimensions of 128 D vector to 3 do you have a compirasion of tSNE (or other dimensionality reduction methods, PCA) to HTM's spatial-pooler? Btw, for images, you could try a biological RetinaEncoder (as a working PR in this repo) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#782 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJA2ECKVPSNNFSC2IIHWVVLRMQP6TANCNFSM4LIFROQQ> .

-- Kind Regards Aamir Mirza

[aamir-gmail]

Hi, Along with TSNE, the 4th dimension is the norm ( magnitude ) of the 128 dim original vectors. That is the one which makes all the difference, so the final SDR as input to SP is, 6000 bits for the original 3D ( TSNE reduced ) vector plus 2000 bits for the norm of the 128 bit. Finally an SP with ( 8000,) and input at (8000,) bit SDR. I was looking for an eye encoder I could not find in the master branch. If you have a link with some examples on how to use it please let me know. Cheers Aamir

…

On Tue, Apr 14, 2020 at 6:52 PM breznak ***@***.***> wrote: thanks for the heads up, Aamir! I have used TSNE to reduce the dimensions of 128 D vector to 3 do you have a compirasion of tSNE (or other dimensionality reduction methods, PCA) to HTM's spatial-pooler? Btw, for images, you could try a biological RetinaEncoder (as a working PR in this repo) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#782 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJA2ECKVPSNNFSC2IIHWVVLRMQP6TANCNFSM4LIFROQQ> .

-- Kind Regards Aamir Mirza

[breznak]

I was looking for an eye encoder I could not find in the master branch. If
you have a link with some examples on how to use it please let me know.

the retina encoder is here, in an almost working state.
#691

Along with TSNE, the 4th dimension is the norm ( magnitude ) of the 128 dim
original vectors. That is the one which makes all the difference

thanks for these findings! If you're interested, we could include this as an image encoder in here.

Yes, I tried PCA as well as UMAP, however, TSNE worded best,

actually, I was asking if you've compared tSNE (+SP) vs raw data (+SP --meaning the SP does all the reduction, and how well it performs?)

[aamir-gmail]

Hi, I have not used the 128D raw data on the advice received that this will be too much for the SP to handle( Numenta discourse ) please see my original post here https://discourse.numenta.org/t/how-many-feature-can-be-concatenated-for-input-to-spatial-poler/7224/15 . encoding one float32 to 1000 bit SDR would require 128000-bit input to SP. If you think SP in HTM core could handle it I am willing to give it a go. Shall I use one RSDE encoder to the entire ( matrix ) 128 D features plus 10K image dataset, or separate RSDE per column? If you wish to use parts or full code for the image encoder them please do so. I have no problem. Kind Regards Aamir Mirza

…

On Wed, Apr 15, 2020 at 12:44 AM breznak ***@***.***> wrote: I was looking for an eye encoder I could not find in the master branch. If you have a link with some examples on how to use it please let me know. the retina encoder is here, in an almost working state. #691 <#691> Along with TSNE, the 4th dimension is the norm ( magnitude ) of the 128 dim original vectors. That is the one which makes all the difference thanks for these findings! If you're interested, we could include this as an image encoder in here. Yes, I tried PCA as well as UMAP, however, TSNE worded best, actually, I was asking if you've compared tSNE (+SP) vs raw data (+SP --meaning the SP does all the reduction, and how well it performs?) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#782 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJA2ECPSGVFCNGDJDMOKSKDRMRZFZANCNFSM4LIFROQQ> .

-- Kind Regards Aamir Mirza