Methods for predicting multiple output fields

Hi folks. As the title says, I’m wondering how to predict multiple output fields from a temporal mmory’s SDRs. For example with the well-known hotgym example, one can use a simple ML regressor to translate from the minicolumn or cell SDRs into a scalar power value, but those SDRs also contain timestamp data. What’s a practical way to simultaneously extract all of that information?

I’ve glanced through the thread here but didn’t gain any insights about the methods used.


I’m wondering would n ML regressors, one for each kind of variables mixed in SDR, perform well in this case? Have you already tried it?

I think the information is still pretty separable, because each output bit:

  • either mixes values of different kinds and then these mixed values are somehow correlated, i.e. loosely interchangeable (= output bit provides almost the same information regarding each of the mixed value)
  • or it doesn’t mix them and they’re separate

So, as I understand, SP is a pretty good encoder in terms of preserving information.


That was my first thought too–I’ll try it out and see what happens.

1 Like

I’m not sure there is a straightforward way to do that.

I think the shortest route to your goal would be to just make multiple copies of the model, one for each field you want to predict. You’d just have the tweak the inferenceArgs in the config file – specifically the predictedField.

There’s a GitHub issue for this too:

It seems to be implemented in htm.core but not NuPIC if I understand, which is believable since NuPIC has been in maintenance mode for some time.