Closed kimtonyhyun closed 4 years ago
I then performed the CPD analysis:
The best r-15 model, trial vector colored by correctness:
The results can be obtained here:
c11m1d13-export.mat
: Contains the exported data tensor X
;c11m1d13-nncpd_r15_run01.mat
: Contains the best-performing r-15 model cpd
and the reconstructed tensor Xest
.I then used eval_fit
to consider the quality of CPD fit against the cells 69, 278, 310 and 497.
Cell 69 (Rsq=0.4315). I sorted the individual trials by residual:
Cell 278 (Rsq=0.5841):
Cell 310 (Rsq=0.6086):
Cell 497 (Rsq=0.5423);
Based on these examples, I don't think the r-15 CPD fit is doing an adequate job fitting the exemplary mixed-selectivity cells in c11m1d13. The reason for this could be that the example neurons have Ca2+ traces with transient dynamics, which is not adequately captured by the CPD model.
Thoughts?
Here is a distribution of single-cell Rsq values for the 27 mixed selectivity cells of c11m1d13:
My rough threshold for a "good fit" is Rsq > 0.7. Out of the 27 cells, only one reaches this value, and the rest are at the bad-to-okay range.
Here is the best-fit example, Cell 477 (Rsq=0.7321). Individual trials sorted by residual:
Here is the corresponding raster plot for that cell. It seems to me that this cell has the least-transient-like trace among the 27:
@ahwillia As a first look at mixed selectivity in the mPFC dataset, I shifted manually through c11m1d13 for cells that showed the phenotype at a single-cell raster level.
As an aside, we may want to consider a different name for these cells in our dataset. I think our phenotype is somewhat distinct from "mixed selectivity" of Fusi et al. in that the different factors are at play in a single cell at different times. (I think this is a good short summary of their definition.) We can discuss more in person.
To follow along, go to __Dropbox/schnitzer_ganguli/c11m1d13__ in Matlab:
DaySummary
instance.mixed_cells
, which is a list of 27 cells (out of 507; around 5% of the population) that I considered to be "mixed selectivity cells" based on their rasters.You can then run
browse_rasters
on the mixed-selectivity subset of cells as:Here are the 27 single-cell rasters:
Among these, I think Cells 69, 278, 310 and 497 have particularly high "fidelity". Here are their detailed single-cell rasters. Feel to speculate on what their interpretations might be :grin::