How important is it for the CNN that the data is perfectly “grid-like”?
I know CNN’s are great for images because the pixels in an image are placed like in a grid with constant distance to each other. This grid can be 1 dimensional as well as for a time series with constant update frequencies.
My data that are a time-series and should thus work with CNNs as well. Unfortunately the data is not precisely gathered with a constant interval. And is thus not perfectly in a grid.
It might look something like this: [6,6,6,6,20,20,20,10,10] etc. where the number is time since last data-sample. So first there are 6 seconds between the data samples, then there is 20 seconds etc.
Do i need to interpolate between my data samples so that it everything is updated with an interval of 6 seconds for my CNN to work? Or will it work just fine using the changing interval?
I don’t seem to be able to find much information on the subject. From deeplearningbook.org under the CNN chapter they mention that its important that the data is “grid like”. However, they don’t comment on if its impossible to use CNNs if the data have irregular update frequencies. Just that its important.
Any insight or links to any information regarding this is very appreciated.