We have released all our commercial HTM algorithm code to the open source community within NuPIC. We use that code in our product. The NuPIC open source community and Numenta continues to maintain and improve that regularly.
Internally though we continue to evolve and expand the ideas towards a full blown cortical framework. Those research ideas are constantly in flux as we tweak and experiment. To go along with that we have a separate experimental codebase that sits on top of NuPIC.
We get a lot of questions about it. As such we wondered whether it is possible to be even more open about that work. Could we release our day to day research code in a public repository? Would people get confused? Would it slow us down?
We discussed these tradeoffs on the NuPIC mailing list. Based on that
discussion, we decided to go ahead and create nupic.research It contains
the code for experimental algorithm work done internally at Numenta.
The code includes prototypes and experiments with different algorithm implementations. This is all temporary, ever-changing experimental code, which poses some challenges.
Hence the following DISCLAIMERS:
- the contents can change without warning or explanation
- the code will change quickly as experiments are discarded and recreated
- it might not change at all for a while
- it could just be plain wrong or buggy for periods of time
- code will not be production-quality and might be embarrassing
- comments and questions about this code may be ignored
- Numenta is under no obligation to properly document or explain this codebase or follow any understandable process
- repository will be read-only to the public
- if we do work with external partners, that work will probably NOT be here
- we might decide at some point to not do our research in the open anymore and instead delete the whole repository
We want to be as transparent as possible, but we also want to move fast with these experiments so the finalized algorithms can be included into NuPIC as soon as they are ready. We really hope this repository is helpful and does not instead create a lot of confusion about what's coming next.
OK, enough caveats. Here are some installation instructions though mostly you are on your own. (Wait, was that another caveat?)
Requirements: the main requirement is nupic. Various individual projects may have other requirements. We don't formally spell these out but two common ones are pandas and plotly.
Install using setup.py like any python project. Since the contents here change often, we highly recommend installing as follows:
python setup.py develop --user
After this you can test by importing from htmresearch:
python
from htmresearch.algorithms.general_temporal_memory import GeneralTemporalMemory
If this works installation probably worked fine. BTW, the above class is a modified version of TemporalMemory that we are currently researching. It includes support for feedback connections (through apical dendrites) and sensorimotor inference (includes efference motor copy as another option input to the layer).