how do we see improvement in decoding with corrs in presence of bad noise-corrs?

[mobeets]

blue dots in top right quadrant: they have positive noise correlations which should mean shuffling HELPS them (i.e., negative score --> bottom right quadrant).

same for red dots in top left quadrant

[mobeets]

one realization: we are calculating noiseCorrAR using spikes minus model predictions, i.e., an estimate at when the signal is zero. but the noise correlation cartoons (averbeck style) are really about noise correlations given the stimulus direction. so it's not too unreasonable, especially given #101, to assume that the noise correlations are different given a strong signal.

so i've added a noiseCorr_avg field which calculates the noise corr for the two stimulus direction conditions, using actual spikes (since this is what we use to decode), and then averages the two.

[mobeets]

here's the new plot, comparing scoreGain to the avg noiseCorr of the two different stim dirs:

so it's basically identical to the original noiseCorrAR:

[mobeets]

also though, we can get p-vals on the noise-corrs and only plot the ones with an avg p-val below, say, 0.10:

still, those weird cells exist! the blue dots in the the top-right quadrant