Wishlist

[breznak]

Unordered, confusing, just ideas-drop list.

• top down compute (so we can get input for given SDR, without Classifier)
• Performance optimizations Optimization for performance #3
• Re-Organize repo structure into more isolated, modular APIs
• move to Python 3
• idealize goals, directions of this fork, publish

[mattheiler]

I would add a wiki with some or all of:

• Road Map
• Developer Notes / Style Guide
• Design Documentation (Google Docs?)
• Known Issues

I also wouldn't mind using a code beautifier and linter and explaining expectations in the developer notes.

[breznak]

+1 for the wiki, It's enabled here and should be editable.
I guess we could start by stealing the Style guides, Design doc etc from numenta's? :)

[mattheiler]

Looks like the style guide link is broken. Not sure where it is.

[mattheiler]

I added a few pages... to be filled in later.

[dkeeney]

For the "SDR" data type I would like to consider a bit array (Int64[32]) for 2048 bits. Wrapper this in a class to add bit get/set and AND/ OR and other operations on the entire array. Taking advantage of the 64 bit word to do some things in parallel. Counting overlap bits could benefit from this. I have not flushed out my ideas entirely but it would be interesting to build a prototype and benchmark it against a sparse array or even an Int32[2048] or Byte[2048] array

[breznak]

the "SDR" data type I would like to consider a bit array (Int64[32]) for 2048 bits. Wrapper this in a class to add bit get/set and AND/ OR and other operations on the entire array. Taking advantage of the 64 bit word to do some things in parallel. Counting overlap bits could benefit from this

Definitely good idea for trying! The SDR datatype should be exactly for this - allowing to quickly experiment with different data-structure for the entire NuPIC/SDR.

There is a library for "compressed bit representation", which I think would fit your use-case. bitset?
http://www.cplusplus.com/reference/bitset/bitset/

It would be good for memory footprint, alternatives are sparse vectors.
For speed I wonder if the CPUs are optimized for float operations better.
All in all, we must try and bench! :)

[ctrl-z-9000-times]

I'd like for the temporal memory to accept predictive inputs (onto distal basal dendrites) from an external source. Right now there is an example of this in numenta/htmresearch. This feature is needed to reproduce some of Numenta's papers.

This feature request now has its own issue: #88 (comment)

[breznak]

I'd like for the temporal memory to accept predictive inputs (onto distal basal dendrites) from an external source

I think this is quite easily doable. Would you like to work on that? Anyway, could you please open an issue for this with link to the PY implementation?
Is it for the ColumnPooler experiments?