I'm new in EEG signal classification. Studying the literature on this topic, I wonder why the EEG signals are divided in epochs, so, instead of classifying the whole signal all at once, we usually classify "windows" of signal, splitting the original one. I searched in the literature for a possible explanation, but I find nothing which answers to my question.
Basically, to perform a spectral analysis. Take a look to this: http://support.ircam.fr/docs/AudioSculpt/3.0/co/Window%20Size.html
I do not know if you refer to some specific approach, but usually, the main EEG classification approaches use frequency analysis (also combined with temporal analysis) to extract the features that will be used to train the classifier or infer the prediction.
- Compute-expense: recordings may be tens if not hundreds of thousands of samples long with multiple channels. One either runs out of memory, or heavily downsamples at layer 1, which loses too much information before the NN can do much with it.
- Data augmentation: for e.g. classification, if you have a 10 minute segment belonging to same class, splitting it in two gives you twice as many training samples
- Feature localization: small kernel length convolutions are generally preferred, which may better extract time-localized information than global - unless one explicitly increases receptive field with more layers and strided convolutions (like in WaveNet).
- Stability with networks that suffer gradient problems for very long sequences, like RNNs. More info doesn't help if a net can't learn from it (see e.g. here).
There's nothing that says we must divide up a signal, however; depending on compute-resources, amount of available data, and architecture design, feeding all at once can outdo divisions (since presumably the network is exposed to more relevant information at once).