Say I use an RNN, whatever is the cell's type, to perform time series classification. It can thus be seen as sequence classification. The time series is split into random, equal size, overlapping window. Once trained the goal is to make real time classification. So, here is tow scenario:
- Each window is a sequence of
nsamples used to train the RNN on
ntime steps. Then, the real-time classification is done with a sliding widow.
In that case: (1) is there any other methods than using a sliding window in real time ? (2) Is there any links that is made between a window and the previous one(s) ?
- From each window is extracted a summary vector.
In that case: do i have to construct sequences of
n summary vector from
n windows, or can I use a one time step RNN which takes the last previous internal state into the first current internal state ?
This post comes from the fact that, from what i saw, there is no temporal links that are made between sequences. RNN's are just trained with randomly picked up sequences. But if tow sequences follows, the second one may be related to the first one. Or the last one may even relate to a distant one.