A　geometrical　representation　of　M-P　model　is　firstly　introduced,　by　which　the　training　problem　of　neural　networks　may　be　transformed　into　the　covering　problem　of　a　point　set.　According　to　this,　the　geometrical　algorithm　of　neural　network　training　is　analyzed.　The　algorithm　may　be　used　for　constructing　very　complicated　classifying　boundary,　but　it　has　higher　time　complexity.　So　a　fuzzy　classifier　based　on　the　combination　of　the　covering　approach　and　fuzzy　set　theory　is　proposed.　The　classifier　can　improve　the　speed　of　training　and　decrease　the　number　of　covering　sphere-neighborhoods,　i.e.,　decrease　the　number　of　hidden　nodes　of　neural　networks.　The　fuzzy　set　based　approach　may　also　provide　multi-choices　for　pattern　recognition　problems　of　large　scale.　Recognition　of　700　handwritten　Chinese　characters　is　used　to　test　the　performance　of　the　approach　and　the　results　are　promising.