RNA sequencing (Tasic at al 2018, see kobak and berens 2019)
FLYEM team
Behavioural data (lots of it)
Variance vs PC dimensions (eigenspectrum) - it follows a power-law decay, top few PC dimensions are not enough.
t-SNE or autoencoders for unsupervised approaches. or the clustering output.
The resulting method from this is Rastermap.
Step 1 - cluster the data with k-means
Step 2 - Rearrange clusters by similarity - sort the matrix of similarities between cluster activity. (keep power-law decay of PC variances) - iterative optimisation (NP HARD)
Step 3 - Upsample cluster centers (lin interp I think)
Step 4 - Assign neurons and sort (done)
Benchmarks with neighbour metrics (local score) and also intermediate and global score of distances. [the idea is neighbours in Nspace vs those in linspace or 3dspace]
Can visualise the behavioural predictions of (behaviour estimate neural activity).
Ahrens lab zebrafish data - can colour neurons in the brain.
Rastermap taking a space filling curve through the data?
Check out the rastermap GUI (how does it load the data?)
For CI imaging they cluster the deconvolved data. (long decay time of the sensor, you can get rid of the decay by deconvolution - increase SNR due to bleedthrough)