HTM is a theory of neocortical function developped by Numenta with an open research philosophy. You can read about it on https://numenta.com/ or participate in the active community at https://discourse.numenta.org/
HTMATCH aims to be a C++ architecture for simulating cortical areas, taking advantage of biologically plausible topologies, and typical range at which cells can connect to their neighbors, to in fact increase performance.
It is a strong design intent to keept the final framework model-agnostic, but allow it to run models abiding to HTM philosophies in particular (hence the 'HTM-Able' designation).
Current proposal of optimized, but "vanilla" HTM implementations is but a first step.
- Optimized version of HTM Spatial Pooler
-
- should work best when compiled for x64 targets, with either MSVC or GCC (requires c++17)
-
- perf reports to be found at https://discourse.numenta.org/t/topologies-in-the-brain-and-how-to-model-them/6783/15
-
- code is a little heavy on preprocessor macros at this point, having devised this scheme allowing to select a particular version and run tests against it. Tried my best to make this hopefully manageable enough.
-
Further optimized versions of HTM Spatial Pooler (yet faster local inhibition)
-
Implementation of HTM Temporal Memory.
- Topological architecture intended to support multi-area computational models