make a default classifier from multiple datasets about suite2p HOT 4 CLOSED

Trying to use Allen datasets for this, one for each CRE line, except EMX for which we will use several (it's a dense label). Should consider adding neurofinder datasets, which are very different because they are short and more noisy.

from suite2p.

Allen datasets are noisy for most CRE lines and include weird pre-processing steps like pixel deconvolution. Will only use some of their EMX-CRE. Investigating the neurofinder datasets now. Also asking Mario for some of his 2x zoom data.

from suite2p.

need to decide which stats to choose, currently: skew, compact, footprint

from suite2p.

Added normalized pixel count as statistic. Default classifier includes a single mesoscope recording with ~6,000 cells and ~6,000 non-cells at low zoom and 2:3 aspect ratio. The trained data was before we brought back connected regions, so some of the non-cells are scattered. Classifier histogram smoothing works as before except the grids are determined from the training data as quantiles rather than pre-defined.

No other statistic helped performance on test data. We tried aspect_ratio, std, kurtosis, mean image correlation, and correlation with neuropil.

from suite2p.