reusing code: Good python library for ML/metrics,...?

[breznak]

From my experiences explaining HTM/CLA to people used to "classical" neural networks, NuPIC is a very innovative idea. We should benefit from the work/bug-fixes others have done in OSS :smile:

In this issue I'd like to track pieces of code we have in common with classical ML methods, and know about good python (c++ counterpart is a +) libraries for machine learning which we could use.

For example, I have found:

• Metrics: ours: https://github.com/numenta/nupic/blob/master/nupic/frameworks/opf/metrics.py lib: http://scikit-learn.org/stable/modules/classes.html#sklearn-metrics-metrics lib2: https://github.com/benhamner/Metrics
• somewhere we use k-Nearest Neighbors

[rhyolight]

Should we add more dependencies to nupic?

I think that is the bigger question. The feedback I've gotten from the community so far has been to keep as few dependencies as possible. I'm interested in what others think ( @subutai @scottpurdy @chetan51 @oxtopus ). This might be a question to ask nupic-hackers.

[scottpurdy]

Our metrics are a very light-weight part of the codebase and probably don't warrant an external dependency. Plus using an external library really limits our control over the code.

The KNN classifier is a more logical piece to find an external version of but we needed more control over the implementation so we implemented our own.

I think we will have a hard time integrating with external implementations in a lot of cases just because you lose the flexibility to modify it as needed.

[breznak]

good point with the tweaks to kNN.

I picked Metrics as they are just there (not subclassed by anyone, etc) - so a good candidate for a library. Also it'd give us more (types) of them for free.

Another candidate on my mind is Swarming - a blackbox optimization algorithm (but that's rather radical for now;) )

[scottpurdy]

@breznak - regarding swarming, see here: #1236