Cerebellar　motor　learning,　including　an　acquisition　of　ocular　following　response　(OFR),　can　be　reproduced　using　mean　firing　rates　(MFRs)　as　the　error　signal　of　climbing　fibers　(CFs).　But　real　neurons　transmit　the　signal　by　spikes,　which　are　discrete　events.　It　is　not　obvious　whether　learning　is　possible　with　discrete　spike　trains.　To　address　this　issue　we　performed　a　computer　simulation　of　cerebellar　learning　using　CF　spikes　-　instead　of　MFR　-　as　the　error　signal　for　Purkinje　Cells　(PCs).　To　generate　the　spikes　we　used　four　spike　generation　models.　And　we　found　that　in　an　OFR　task　with　a　constant　visual　velocity,　learning　was　successful　with　stochastic　models,　but　not　in　the　deterministic　models.　©　2005　IEEE.