Loose boundaries and strange undifferentiated soup

Eyewire Questions
#1

From time to time I am getting into situations where there is very hard to decide what is part of what. Most of the time its well formed “pipe” going into a sort of strange undifferentiated “balloons” (cloud-like, not really circular formations made up of even smaller badly differentiated “granules”). Not having an xray vision, at first I try to find boundaries of that “balloon”. Sometimes it works and makes sense (on 3D display mostly) but many times it just starts meaninglessly filling-up screen with amorphous formations.

In short, sometimes I have serious problems with boundaries.
What am I to do? Should I strictly follow visually observable limits independent of what I get? Can we get some exemplary, practical images? Our neural networks need training too.

Here is a somewhat less stark example:




Stubbornly coloring without violating boundaries, I end up with something like this:
Notice that “cloud” which has a shape but is made of quite ugly chunks. Unfinished.

gap3.png693×587 266 KB
 
#2

Hiya Sidran,

Sometimes the lines which separate two neurons are hard to see, sometimes they’re completely invisible. The tissue is stained so that cell walls appear darker when imaged with an electron microscope, but the stain doesn’t always reach all of the cell walls.

The AI can get confused in these areas just like you and me.


To me, it looks like the area above the red dashed line here should be included, but what’s below it is separate. Unfortunately, when you tried adding this, it selected both pieces (the AI merged these two pieces together). In this case I would have not selected the piece.

Now let’s assume you didn’t notice, and you kept adding things (as you did in the picture). After a while, it starts to look like the pieces you’ve selected are “hugging” the original piece of neuron, not really growing out of it. After you’ve selected so much, it should be more obvious that the two pieces are separate, so you should deselect all of the extraneous pieces.

Does this help?

#3

Not really, but I appreciate your goodwill. I am not sure how far can we get with this if only a few dozen gray areas stay inadequately connected. Through error or more importantly through lack of knowledge. Some training image gallery (3D view of app itself) might be good to have. A lot of correctly connected parts classified by their type (dendrites, axons… everything that might be encountered) of varying difficulties allowing people to react more confidently. Variety and difficulty is important as well as having images from app itself not some book or unrelated illustration.

This ceases to be a question and is now a suggestion.
#4

sidran I see what you mean, I also got this frome time to time. 


As echo says, to solve the problem I have the following approach.

If the resulting (new, light blue) shape looks like a branch (e.g. it grows from the initial, dark blue shape), it is safe to assume it is the same neuron so I keep it coloured.

If the resulting (light blue) shape seems to start elsewhere and just passes by (hugs) the initial shape (dark blue) I guess it is part of another neuron and I let it be (uncoloured)
#5

That’s exactly right Frabuleuse.  Thanks!

#6
sidran I see what you mean, I also got this frome time to time. 

As echo says, to solve the problem I have the following approach.

If the resulting (new, light blue) shape looks like a branch (e.g. it grows from the initial, dark blue shape), it is safe to assume it is the same neuron so I keep it coloured.

If the resulting (light blue) shape seems to start elsewhere and just passes by (hugs) the initial shape (dark blue) I guess it is part of another neuron and I let it be (uncoloured)

I do try that approach… and it does miracles. Biggest problem is when I encounter a sort of blobs which seem to go nowhere or do not have any specific shape. We’ll keep trying, thanks for your suggestion.

#7

Sidran, if you post a photo of the next blob type thing you see I can help you figure out where it’s likely to belong.  Something to keep in mind is that everything in the cube has to belong to something else in the cube–there are no blobs just floating in space in the brain, everything is attached to something else.

#8
Sidran, if you post a photo of the next blob type thing you see I can help you figure out where it's likely to belong.  Something to keep in mind is that everything in the cube has to belong to something else in the cube--there are no blobs just floating in space in the brain, everything is attached to something else.

The fact that everything is attached to something might be obvious to you but was quite informative and reassuring for me. I thought there are some sporadic, unattached “pockets” of fluid/goo/whatever floating in between. Being able to access neighboring (at least 6) cubes would possibly help in tight situations.

Lately I am doing just blitz sessions… hopefully will return to help some more.
#9

Very true that accessing neigboring cubes would be useful in demystifying difficult spots–eventually we hope to offer something like that (or perhaps bigger cubes).  We are currently working on a feature that would allow players to see the cube from a different axis.  Hopefully that’d help as well.

#10

One issue with the way Eyewire is gamified is that we’re rewarded for being generous, instead of conservative, in judging these ambiguous boundaries (AFAICT, more clicks = more points).  At the least, the tutorial should train us to be, or explicitly state that we should be, conservative.

#11

We don’t want people to be too conservative.  So far it appears that people are missing more things than they are incorrectly adding.    You don’t get more points by clicking more.  Your results are compared to the several other people who have done the same cube, your points come from how closely your results match theirs, and the amount of time you spent on the cube.