Temporal resolution vs. precision, pre- and post-convolution

[@jyanguas3251]

Great video! One question: Why does filtering impose a loss of temporal precision? Thank you

[@mikexcohen1]

Each time point in a filtered signal is defined as the weighted average of neighboring time points, where the weights are the filter kernel. So each time point is influenced by its neighbors.

[@tobi3497]

I was also thinking about the implications of maintaining 100% time resolution. I thought that saving the image using compression codecs like jpeg would help, as they are designed to compress periodic data (I think using FFT/Wavelets internally). I've not actually experimented with this yet. Essentially, I feel like you'd get better results relying on an image compression algorithm saving at a smaller resolution, as it deals with downsampling and frequency compression - all in one fast algorithm - rather than downsampling yourself. As you stated, the goal is for human analysis of the downsampled spectrograms to derive the same results to the original, which is what image compression algorithms are designed to do.

[@mikexcohen1]

That's a really interesting suggestion -- using image compression algorithms as feature-extraction for time series data. I never made the connection and I haven't come across it (which of course doesn't mean that it isn't already a researched topic). From a signal-processing perspective, that sounds like it could be promising. From a neuroscience perspective, I'd be a bit cautious about it, because I like having some physiological interpretation of the data, and it's possible that the image compression algorithms are mixing signals in weird ways or preserving noise at the expense of signal.