Query about opf model

I was looking at the opf model (link). I had this query: Basically, the TM predicts the next set of neurons that is going to fire. So, in the link above, during training, we are sending SDRs in which every SDR is a combination of timestamp and consumption value. So, the next prediction by the TM is going to a set of neurons (that too an SDR) which will even be a combination of some timestamp (time of day, weekday or not) and predicted consumption value…right?

If thats so, how are we getting the predicted consumption value and how are we inferring what timestamp does it correspond to?

And can someone please explain about multiStepBestPredictions. I dont get the point of best prediction. I thought there was only one prediction at any point in time by TM. Is it related to the case where there might be multiple predictions (like eats word in the HTM school episode). If so, then it means multiStepBestPredictions make sense only incase of ambiguous predictions (when there can be more than one prediction)?

And whats 5 step prediction? Is it like the model predicts for the next timestep, gives that as input to get the prediction for next timestep and so for 5 times and gives the output?

Is 1-step and 5-step even comparable?

And can I know whats aggregation in the params.yaml file?

1 Like

NuPIC uses a standard ML classifier to decode the neurons into a value matching the input.

At any point, the TM’s predictive cells could contain multiple predictions. The best prediction is the one with the highest probability, but others are included in case you want to parse through them.

You can set up the model to predict as many steps in the future as you want. The further out you go, the worse the predictions will be. If you choose 1,5 then you’ll get predictions for 1 and 5 steps out. You can add more, but it takes longer to process.

It is about time aggregation, and I never use it. :man_shrugging: I prefer to aggregate my own data.

1 Like

Thanks for the reply.

But still I have this doubt. The predicted SDR is a combination of [predicted timeofday + predicted weekend + predicted consumption]…right? How does the classifier decode the consumption value alone? I couldnt find that explanation in the classifier video shared earlier.

1 Like