This repository contains the code for experimental algorithm work done internally at Numenta. A description of that research is available here.

We have released all our commercial HTM algorithm code to the open source community within NuPIC. The NuPIC open source community continues to maintain and improve that regularly (see https://discourse.numenta.org for discussions on that codebase. Internally we continue to evolve our theory towards a full blown cortical framework.

We get a lot of questions about it and 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 decided to go ahead and create htmresearch. It contains experimental algorithm code done internally at Numenta. The code includes prototypes and experiments with different algorithm implementations. It sits on top of NuPIC and requires you have NuPIC installed.

Our research ideas are constantly in flux as we tweak and experiment. This is all temporary, ever-changing experimental code, which poses some challenges. Hence the following DISCLAIMERS:

• the code can change quickly and without warning as experiments are discarded and recreated
• code will not be production-quality, buggy, or well documented
• 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?)

pip install nupic htmresearch

Requirements:

• nupic and nupic.core
  • pip install nupic --user should be sufficient
• htmresearch-core
  • To install, follow the instructions in the htmresearch-core README.
• 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.apical_tiebreak_temporal_memory import ApicalTiebreakPairMemory

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 external distal basal connections.

You can perform a more thorough test by running the test script from the repository root:

%> ./run_tests.sh 

If you encounter this error after installing pytorch on Mac OSX:

ImportError: dlopen(~/Library/Python/2.7/lib/python/site-packages/torch/_C.so, 9): Library not loaded: @rpath/libc++.1.dylib
    Referenced from: ~/Library/Python/2.7/lib/python/site-packages/torch/_C.so
    Reason: image not found

Use the following command to fix your pytorch installation:

 install_name_tool -change  @rpath/libc++.1.dylib /usr/lib/libc++.1.dylib  ~/Library/Python/2.7/lib/python/site-packages/torch/_C.so

Some of our old research code and experiments are archived in the following repository:

• https://github.com/numenta-archive/htmresearch