@mo_lotfollahi @david_sontag @RBoiarsky Their clustering abilities (even with fine-tuning) had a lot of issues beating PCA here https://t.co/et8dahjYlz I also remember seeing a paper of yours on cell type classification (whose conclusion was that the "gro
@meltager @biorxivpreprint If I may toot my own horn (poor taste I know), you may be interested in a similar work we published in 2020 https://t.co/et8dahjqw1. We were less extensive on the model architectures (juste vanilla VAEs with different losses), b
@trizzlor @anshulkundaje As for parameter benchmarks, this is very very compute intensive. For our data integration benchmark we used defaults/heuristics for all parameters, but needed 55k CPU hours for only 1 round of revisions. The only parameter benchma
@GreeneScientist @autobencoder (shameless plug) Regarding the "Author Degrees of Freedom" you can see most recent methods fail to reliably outperform PCA, as properly tuning them is incredibly hard (https://t.co/et8dahjqw1)