As I understand, confirming a dendrite is done by comparing the work of 5 tracers.
This is justified for complex cells but some types have few long straight dendrites with rare branchings (see image from OSAC ).
Once they are traced - let’s say twice - most of it seems clear.
Aftwerwards it is rather boring to get in 95% of cases cubes with usually already quite complete “sticks”
while rarely there is a “new job” with a terminal end for eventual prolongation.
I assume that the “safety margin” for such cells could be lowered to 4 or even 3 repeats
This should speed things up and make it more attractive to work on them.
Perhaps this relaxation could be restricted to the distal range relative to the cell body.
Hey peterahn,
This is an interesting discussion topic. We would love it if we could get our coverage down to 3 or 4. They are just rough numbers, but our current coverage is closer to 20. And you know what? At 20 coverage, we are still missing branches! This is especially true of the very difficult Starburst Amacrine Cells (SACs) that we’ve been working on recently, but it’s also true of some of the other cells that we’ve worked on which aren’t quite so difficult.
The core issue is the following. At some random points along these long straight branches there are other branches sprouting off them. In order to find all the branches, we need to look at each piece of each branch in great detail. At the moment, we have no real way of deciding that one chunk is more likely to have another branch sticking out of it than another other than having enough users agree that something should be included. At the moment, the way that we do it is we include a branch as soon as 3-4 people agree that something should be included. As I said before, sometimes out of 20+ people, we can’t even find 4 who get it right.
The other issue, which you might not have considered is that not all of our players are created equal. Some of them are still learning how to play, and we have to account for the possibility that we might have some vandals who include random garbage. These things prevent us from lowering the amount of coincidence that we need to be too low, otherwise we could create a lot of mergers.
There are a couple ways of dealing with this. The first is better player training. If our players are more likely to find the right pieces, then we get above threshold faster and we need less coverage. This might also be improved by a better incentive structure. Longer term, we are hoping that the AI can save us in this regard. What if, instead of having low coverage on the “easy” ones, we had no coverage at all. If the AI could tell us both the right answer and how confident it was in that answer, then we could just serve up the difficult ones for our players. Something tells me this would be much more fun by your definitions. It would also drastically speed up our progress. Unfortunately, we don’t really have anything like this coming down the pipe yet, but we do have a bunch of smart scientists working in the lab on this sort of thing. Hopefully, this is one of the ways that we’ll be able to make Eyewire more fun and more productive in the future.
We come back to the scoring system. A few times I’ve found a ‘hidden’ branch and whilst I’ve still marked and submitted it I think “Here’s another 20 cube”. The AI possibly only knows you’ve marked something different to everyone else so far and marks you down. Don’t know if its possible but retrospective scoring would help.
+++ on retro scoring.
Yeah, I think retro scoring may be the way forward. We’re trying to work out how to do it well.