In NuPIC when you want to get the predicted state, what bits do you use ?
Active, Predicted OR union of both ? or some other set ?
This might help answer your question:
Also, @ycui has been working on an SDR Classifier to replace the old CLAClassifier. I’m not sure what the state of that is at this point, hopefully it can get into NuPIC soon.
@mraptor I have been doing some classification experiments with HTM. If you are using the activity of TM as input for a classifier, I suggest to try predicted-active cells first. Because this representation captures both the sequence context and current feedforward input.
@rhyolight The SDRClassifier is included in NuPIC and the SDRClassifierRegion is almost done. After this gets merged one can use SDRClassifierRegion instead of CLAClassifierRegion.
2 Likes
by this do you mean : Union(predicted,active) ?
I mean predicted-active[t] = Union(predicted[t-1], active[t]) (Active cells that are predicted on a previous time step)
1 Like
So @ycui, the parameter patternNZ of SDRClassifier compute function should receive predictedActiveCells output from temporalPoolerRegion, instead of activeCells or bottomUpOut, that is the default output?
When the SDRClassifierRegion be finished, which output will be the default one?
@ruanjsb My suggestion is based on some preliminary experiments I still exploring, and it only applies to sequence classification problems (not online streaming prediction). I don’t think we will change the default use of activeCells. I can see problems in using just the predictedActiveCells for online prediction: at the beginning of the learning there may not be any predictedActiveCells at all. It is more appropriate to use activeCells as input.