Feedback Daniil

[richelbilderbeek]

From a talk with Daniil's notes (eduPrint_scan_2022-04-05-13-09-00.pdf):

• Compare with other methods, e.g.

  • PCA Compare with regular prediction: PLINK PCA and fit linearly #34
  • UMAP Compare with regular prediction: UMAP and linear regression #36

• Indeed, one should use data of sufficient complexity

• Do a SNP injection, to see if NN has learned genotypes #38 Use a neural network that is trained on a real dataset, then inject SNPs and phenotypes. In that way, you should detect if the neural network pick it up yes/no

• Compare with regular prediction: latent layer and linear regression #37: The neural network for the quantitative trait prediction may not be better than a linear regression, especially if one uses only two dimensions. To find out: use the points from the latent layer and let other methods (among other, linear regression) use it as input. In that way, the NN for QT prediction can be valuated correctly

[richelbilderbeek]

As this Issue has been split into many, close this one :-)