We retrained a popular connectome-constrained model of the Drosophila motion pathway from scratch, twice. Task performance was stable across runs. But whether task-error selection picked the biologically-matched solution was not.
A cautionary note on biological inference from connectome-constrained networks trained in supervised settings: model selection based on task performance alone does not reliably recover the true biological circuits or underlying mechanismsView quoted post on X
Led by @ckaraneen biorxiv.org/content/10.648…