Normalize the length of EEG signals?

[liutianlin0121]

I am doing some experiments on classification with Conceptors. I wonder if it makes sense (in a neural science perspective) to normalize the length of EEG signals? If it is legal, then the data set could potentially be handier to train on Conceptors. Many classification tasks (like "vowel classification tasks") do this "signal length equalization" in pre-processing stage, but I am not sure this could be done on EEG signals. Have you encountered any EEG literature do something similar?

I wonder if Prof. Godde commented on your pre-processed EEG signals? Does him (or you) feel that there is still much space on further polishing the pre-processed data? I thought that, if the pre-processed data has already been recognized decent, I will focus on trying different classification algorithms, without worrying too much about quality of the data.

[angerhang]

For the EEG signals, one will still need to apply the notch filter at 50 Hz. Secondly, normalization should not be done on EEG signals of different subjects if possible.

The general recording length should be the same, so it is ok it just discard the last bits of the recordings to the shortest EEG for normalization purpose.

But before doing any classification, one really needs to focus on the methodology because successfully using ESNs or connectors to classify EEG is nothing non-trivial. The real contribution is the methodology to using NN for hypothesis testing. For this, you will need to talk to Godde on what is the hypothesis that one needs to test and then one needs to think about how to test that.

[liutianlin0121]

Thanks for your responses. About the length normalization, obviously, we could (i) discard first bits of the EEG signals to the shortest EEG signal, (ii) discard last bits of the EEG signals to the shortest EEG signal. Why do you think the latter is preferable?

[angerhang]

Either should be fine, as the resting-state dynamics should be consistent in a recording session.