1st
Jul 2016
Jeremy Freeman
@freeman-lab
Jul 01 2016 03:28
Jeremy Freeman
@freeman-lab
Jul 01 2016 03:35
curious what people think
code is all matlab unfortunately
and they don't show any real ground truth
but maybe some interesting algorithmic ideas
the biggest computational speed up seems to from the SVD (section 5)
Davis Bennett
@d-v-b
Jul 01 2016 03:52
@freeman-lab do you have an intuition for how the SVD trick works? There's some parametrization I don't understand (like why they bin time to get ~10k timepoints -- is this just to fit under a memory ceiling?)
Jeremy Freeman
@freeman-lab
Jul 01 2016 03:58
so the SVD thing works like this
as with many approaches, the setup is an optimization where you are minimizing the squared error between the real data and the model-generated data
and you find parameters that minimize error
if you then approximate the cost function by replacing the data part with the data including only the top k dimensions
then you might as well replace the model part in the same way
so you just end up fitting parameters to a much, much smaller data set
and yeah the binning is probably just to speed up their initial SVD calculation
my bet is, with the SVD, all the model fitting is super fast
Jeremy Freeman
@freeman-lab
Jul 01 2016 04:03
not unlike the distributed ICA we do, where as you know the SVD step is the slowest
Davis Bennett
@d-v-b
Jul 01 2016 04:09
I wonder how many visual-cortex-specific parameters are baked into this thing
"The basis functions B do not need to be fit and are fixed as a
set of isotropic 2d raised cosine functions. The raised cosines tile the full FOV, with inter-center
spacing of ten times the diameter of a cell."
Jeremy Freeman
@freeman-lab
Jul 01 2016 04:09
probably a lot!
Davis Bennett
@d-v-b
Jul 01 2016 04:11
on a scale between "general tool for lots of imaging datasets" and "something that works for one person", my sense is that this is closer to the second pole
Jeremy Freeman
@freeman-lab
Jul 01 2016 04:11
sort of my sense too
we can at least run it on all the neurofinder datasets!
i actually thought @marius10p was doing that
Davis Bennett
@d-v-b
Jul 01 2016 04:14
the simulated ground truth approach could be pretty useful for other methods, e.g. mika's
I'll forward this to him
Jeremy Freeman
@freeman-lab
Jul 01 2016 04:14
how does it work? it's based on hand labeling?
then what's the simulated part?
i still think mika should run his thing on the neurofinder data first :) but yes, definitely good to share with him
Davis Bennett
@d-v-b
Jul 01 2016 04:16
hm i'm not actually sure where the simulated part comes in
"we created a simulated ground truth benchmark. We first ran the pipeline on a single experiment and manually curated the results.We then chose a small subset of cells (20 out of 300) and added their activity to a neighboring spatial location of the movie, which did not overlap significantly with the rest of the ROIs (Fig.
5a-b). We then re-ran the pipeline and compared the results with the 20 known cell locations.
We re-ran this test 43 times, choosing a different set of cells every time. The analysis shows
that almost all ground truth ROIs were correctly identified (Fig. 5c)"
Jeremy Freeman
@freeman-lab
Jul 01 2016 04:17
oh huh
so it's not asking what fraction of a bunch of manually defined ROIs were found
more like a self-consistency check
like, if you simulate from the model, and then run identification on the simulated results
except it's not quite that, because instead of simulating from the model
they're simulating from the data conditional on the model having identified that part of the data as a cell
:confused:
Davis Bennett
@d-v-b
Jul 01 2016 04:20
I'm not sure how to interpret this measure if we don't know how the 20 cells were picked in the first place
Jeremy Freeman
@freeman-lab
Jul 01 2016 04:20
yeah
Nicholas Sofroniew
@sofroniewn
Jul 01 2016 13:02
i think it needs a red nuclear marker