Suppose I want to train a RBM (or even a DBN architecture) and then fine-tune the parameter training a Feedforward NN. In my case the dataset is composed of time series, so in principle there is a temporal dependency between observation.
I use to apply ML techniques which involve LSTM module, so when it comes to split the dataset into the training and the test part this leads to divide it looking at the timeline.
However, It seems that a simple RBM (but also a MLP) is not able to capture the temporal dependencies and the sample are provided as input one at a time. Is therefore reasonable to shuffle a dataset composed of sequences, as it is usually done for other kind of data, as images?
I have a looked for some practical reference about it, but even Hinton's guide gives no idea about this problem.