[Comp-neuro] how can we improve model discovery?