HTM tasks, capabilities - What you can do

[breznak]

Review of encoders' coverage of sensory modalities #259
Here we look at possibilities of HTM networks:

Tasks:

• prediction
  • multi-step ahead prediction:
    • TODO: instead of current "for N-steps ahead prediction, look at N-1-th classifier in predictor and feed it TM(time)", instead do N steps of TM(time+1).predict. Then classify the TM(TM(time+N-i+1)).
• classification: MNIST MNIST example #242
  • advanced vision tasks (ie Retina encoder Retina encoder: biological encoder for vision WIP #691 )
• anomaly detection AD, see Numenta Anomaly Benchmark #205
• (sub) sequence classification - important TODO (using CP, etc...)
• multi-sensory integration - ie by Thalamus Port of thalamus code from htmresearch #676
• sensory-motor integration, ie Htmresearch port cleanup #672

[breznak]

FYI @ctrl-z-9000-times @fcr if you find other category or have more details, I'd like to keep collecting the info here.

[Finnfj]

Hello,
Just wanted to ask what became of experiments with

TODO: instead of current "for N-steps ahead prediction, look at N-1-th classifier in predictor and feed it TM(time)", instead do N steps of TM(time+1).predict. Then classify the TM(TM(time+N-i+1)).

I'm currently using prediction the first time with HTM and thought that the predictor would have to work as you described it, based on the way it is used.
I am not too familiar with classification too, but I imagine this proposed method could maybe improve prediction?