Separating sequences of inputs


I have data that occurs in limited sequences. There are about 8 to 12 inputs per sequence.
What I want to avoid, is that Nupic bridges and learns from the transition from the last input of one sequence, to the first input of the next sequence.

I know it is possible to turn learning off, but I do not think it would solve the problem to simply turn off learning for that transitory step.
Maybe I can turn off all predictive activations in the network right before the first input of a sequence. And to completely avoid possible learning, I think I also need to turn off bursting caused by the first input of a sequence.

Do you think this would achieve my end, and is it possible ?
Or do you think that because of the likely randomness of the transition between the last input of a sequence and the first of the next, the network will not be aversely affected by the wrong patterns it tries to pick up on ?

Thanks, any help is appreciated!

1 Like

What you want is a “reset”. This tells the TM that the current sequence has ended and an new one begins with the next data input. If you are using the OPF, you can call model.resetSequenceStates(). If you are using the temporal memory directly, call tm.reset().

Aha, that solves things more easily than I could imagine.
Thank you, rhyolight!

1 Like