how_can_we_be_so_dense - How to modify to generate binary encodings?

[kyoungrok0517]

Hi, I'm interested in your HTM Spatial Pooler and the paper in the title.

Here's the question: the SparseNet introduced in the paper generates sparse scalar vectors. But the original HTM generates binary vectors. Can I modify the SparseNet to do the same??

[subutai]

the SparseNet introduced in the paper generates sparse scalar vectors. But the original HTM generates binary vectors. Can I modify the SparseNet to do the same??

Hi @kyoungrok0517 - thanks for the note. You are right about the previous HTM algorithms. Those ones used a Hebbian learning algorithm. In this paper we use backpropagation, so scalar vectors are a requirement.

[kyoungrok0517]

Hi, thanks for the response. I'd be much appreciated if you answer one more question.

• What would be the best way to generate binary sparse vectors using SparseNet? Will it be enough to just make non-zero elements to 1 and the rest 0?

[uhfai]

Based on the Sparsenet paper, the sparse matrix encodings are real values and continuous. So no, one and zero encoding is insufficient.

See paper at:

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.391.4486&rep=rep1&type=pdf

Let me know if you have any questions.

Happy holidays.

Bart

On Wed, Dec 18, 2019 at 7:54 AM Kyoungrok Jang notifications@github.com wrote:

Hi, thanks for the response. I'd be much appreciated if you answer one more question.

• What would be the best way to generate binary sparse vectors using SparseNet? Will it be enough to just make non-zero elements to 1 and the rest 0?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/numenta/htmpapers/issues/31?email_source=notifications&email_token=AIIX324F4AKIPDIXQ3ZLAITQZIMOPA5CNFSM4JZIYG2KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEHGAORY#issuecomment-567019335, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIIX323D7NGOUB3B62SCZIDQZIMOPANCNFSM4JZIYG2A .

[kyoungrok0517]

Thanks for pointing :)

2019년 12월 18일 (수) 오후 11:46, uhfai notifications@github.com님이 작성:

Based on the Sparsenet paper, the sparse matrix encodings are real values and continuous. So no, one and zero encoding is insufficient.

See paper at:

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.391.4486&rep=rep1&type=pdf

Let me know if you have any questions.

Happy holidays.

Bart

On Wed, Dec 18, 2019 at 7:54 AM Kyoungrok Jang notifications@github.com wrote:

Hi, thanks for the response. I'd be much appreciated if you answer one more question.

• What would be the best way to generate binary sparse vectors using SparseNet? Will it be enough to just make non-zero elements to 1 and the rest 0?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub < https://github.com/numenta/htmpapers/issues/31?email_source=notifications&email_token=AIIX324F4AKIPDIXQ3ZLAITQZIMOPA5CNFSM4JZIYG2KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEHGAORY#issuecomment-567019335 , or unsubscribe < https://github.com/notifications/unsubscribe-auth/AIIX323D7NGOUB3B62SCZIDQZIMOPANCNFSM4JZIYG2A

.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/numenta/htmpapers/issues/31?email_source=notifications&email_token=AAIAZ7AK2JF7GO34PKRZPO3QZIZS3A5CNFSM4JZIYG2KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEHGK2BY#issuecomment-567061767, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAIAZ7GEXMYY2L7TKSRCWA3QZIZS3ANCNFSM4JZIYG2A .