Extended Data Fig. 1: Comparison of different choices of discretization during training on the zebrafish atlas.
From: STORIES: learning cell fate landscapes from spatial transcriptomics using optimal transport
All methods are run for 10 different initialization seeds. (A) Training and validation losses along iterations and runtime, for the linear method (blue) and a version with 10 forward steps (green), as in [Hashimoto et al., 2016]; (B) Training and validation losses along iterations and runtime, for the linear method (blue) and a version without teacher forcing (orange), as in [Hashimoto et al., 2016]; (C) Training and validation losses along iterations and runtime, for the linear method (blue) and a version with an ICNN implicit step (red), as in [Bunne et al., 2022].
