battaglia01
commented
8 years ago - Improve consistency of thresholding tests for spikes/noise:
- Filtering: globally rescales data so that max(abs) is 1. Should rescale each channel separately, to try to ensure noise has equal variance.
- Whitening: compares cross-channel L2-norm (CCN) to noise_threshold. Probably better to compute L2-norm over a few time bins, and then Lp-norm across channels, to allow for spike waveforms that live on only one or a few channels.
- Clustering: originally used 4*std(CCN) to grab spike segments (??). EPS changed to a constant value of 6, relative to noise stdev (=1). Preferable: User should set a tolerable error rate (miss+fa), provide a prior on firing rates, and theshold should be chosen using an SDT calculation! T= (log(1-pr)-log(pr) + a^2/2)/a, w/ pr=prob of spike in dt, a=spike amplitude eg: for error rate of 0.1%, with a prior of 1e-3, this is about 8.2
- Partitioning (snippets): uses yet another threshold, and additionally relies on min_separation_length and min/max snippet length parameters. This should probably be again, an L2 norm over a few times bins, then combined in an Lp norm over channels.