ANN: Three-nsitions


I created Threensitions. You can read about it here :

What is three-nsition ? Glad you asked :slight_smile:
In two words it is a way to do prediction of Time Series.
Normally you would use Markov chains or Autoregressive models to do that.
Three-nstion is sort of Markov chain with a twist.

What do you think… ?

Q: How does HTM handles zero-freqs problem.
I know that bursting makes a new “state”, but I don’t think it predicts a state which it has not seen yet.

The zero-frequency problem as the names implies is what
do we do in cases where we did not collect statistics yet i.e. new
state show up for the first time. How do we predict them ?
How do you predict something you have not seen yet ?

I’d say it predicts the union of the next step in all learnt sequences involving that particular state.