Alexander Lavin

FWIW I tried an approach where I started the run-lengths matrix with a list `[[]]` and would append rows and columns for each subsequent timestep. This got to be very...

This may work well; it's not clear to me before running it on some datasets. I've implemented a solution as an anomaly detection algorithm in [NAB](https://github.com/numenta/NAB), and will link it...

The online algorithm is implemented [here](https://github.com/numenta/NAB/blob/master/nab/detectors/bayes_changept/bayes_changept_detector.py) as an anomaly detector in NAB.

Apologies for the delay. Yes it looks like a vestigial variable. It's also left over w/o purpose in the SP unit tests. We should clean it up. This also makes...

@subutai relying on numpy.argsort for tie-breaking implies we'll always give preference to higher-indexed columns; if columns 4 and 400 have the same overlap often, column 400 will preferentially win out...

I still don't agree. Yes these are two sides of the same coin (ensuring determinism), but relying on argsort will always prefer higher-indexed columns. In a 2000 column region, you'll...

`id()` does not suffice because it is not repeatable for a given seed; i.e. we want the same random sample to be returned each time we seed with the same...

No I get 70%. I ran again without using my cache and get the same results. I notice there's a difference between hello_classification.py and the simple_labels.py implementation: doc 2 uses...

> > The tests w/ TM (sensor_tm_simple_tp.json config) fail on saving the model in executeModelLifecycle() of htmresearch/support/nlp_model_test_helpers.py. > > @BoltzmannBrain What is the exact command you used? This config file...

> and got 66.2% Okay it looks like I have something locally that is to blame. Just to be sure, what nupic and nupic.core are you using? I'm on nupic...