training mode

Was talking with @kdharris101 about an idea for how to determine whether people are doing the manual stage correctly. It's easy:

User turns on a setting to go into training/testing mode, during which a settable percentage (say 10%) of decisions they make will be about fake clusters (the rest are the real clusters they're trying to sort). So 10% of the time, instead of giving "best unsorted", it gives a fake cluster generated in one of two ways:

1. Inserting a fake cluster and splitting it. In this case, the user is supposed to correctly merge it with its other half. Could split it along peak amplitude, depth on the probe, direction of greatest variance in feature space, or time (for example).
2. Merging two fake clusters that are not too dissimilar from each other and inserting this. The user is supposed to split it. Could parametrically vary how similar are the two clusters that are merged and should be split.

When the user gets finished and makes a decision about this cluster (in the normal way), the program tells them whether they are right or wrong, what they were supposed to do, and gives them the option to try again. If they don't try again, it removes the fake clusters and continues on as before.

It makes a log of the tests proposed and decisions made.

This is useful at three levels:

1. For checking that I myself [i.e. the serious user] is really doing things all right, and for giving myself a sense of how accurate I can be with this stage.
2. For checking up on technicians who might be assigned to learn and do sorting and, similarly, for training new users.
3. As a first step in the direction of developing a crowd-sourcing platform for spike sorting. If that's something one might want to do.