Skip to content

Neural data analysis#

Before starting#

Need to take into account the considerations from

  1. Does the brain care about averages?
  2. Nonsense correlations (see practical example at )
  3. https://elifesciences.org/articles/71969
  4. https://psyarxiv.com/tbmcg/

Stats#

  • q-values (FDR adjusted p-values for many tests).

Repositories#

Pairwise correlations#

  • Just compute a bunch of these, e.g. as in Stringer et al Science 2019.
  • It is hard to interpret though, and feels as though it does not make full use of the population available.

Sorted raster plots#

A common method of activity visualisation is to perform raster plots with sorted neurons.
However: note that the chosen sorting method greatly changes how the data appears.
For instance: A PCA sorting can make data appear quite uniform, while a manifold embedding can make the data look much more chaotic.
This is further emphasised with random sorting.

Shared variance component analysis#

  • Asymptotically unbiased lower-bound estimate for the amount of a neural population's variance reliably encoding a latent signal.
  • Seems to require two populations.
  • Implemented in Stringer et al. Science 2019 but I should look for another place too.
  • Needs thousands of cells recorded per area - so tough without CI TBH.

Peer prediction analysis#

  • Predict acitvity of one neuron from the others.
  • To contrast with SVCA, "SVCA finds the dimensions of activity in a large population that can be most reliably predicted from a held-out set of neurons" Stringer et al. 2019.
  • Can use lower (much) N cells as a result.

gLARA - Group Latent Auto-Regressive Analysis#

  • Instead of analysing direct connections from neurons, latent variables are estimated for different groups, and then these are related to eachother.
  • Similar to pCCA - probabilistic canonical correlations.

ANNs#

  • Usually either goal directed or data driven. See cunningham 2017 for more.
  • Effective in areas such as PMC and retinal ganglion, but not sure on multi-region ANNs.

To check#