If we check the perturbation disentanglement to make sure we're not accidentally training an autoencoder, then we can run a regular test-set reconstruction loss.
Otherwise we will not be able to discern models that perform well averaged-across-cells (eg by predicting the mean) from models that perform well on each individual cell.
If we check the perturbation disentanglement to make sure we're not accidentally training an autoencoder, then we can run a regular test-set reconstruction loss.
Otherwise we will not be able to discern models that perform well averaged-across-cells (eg by predicting the mean) from models that perform well on each individual cell.