Training　of　convolution　neural　network　(CNN)　is　a　problem　of　global　optimization.　We　hypothesize　that　the　more　smooth　and　optimize　the　training　of　CNN　goes,　the　more　efficient　the　end　rsult　becomes.　Therefore,　in　this　short　paper,　we　propose　a　modified　resilient　backpropagation　(MRPROP)　algorithm　to　improve　the　convergence　and　efficiency　of　CNN,　in　which　global　best　concept　is　introduced　in　weight　updating　criteria,　to　allow　the　training　algorithm　of　CNN　to　optimize　its　weights　more　swiftly　and　precisely　to　find　a　good　solution.　Experimental　results　demonstrate　that　MRPROP　outperforms　previous　benchmark　algorithms　and　helps　in　improving　training　speed　and　classification　accuracy　on　a　public　face　and　skin　dataset　[1]　up　to　4X　(four　times)　and　2%　respectively.　In　RPROP　[2],　the　change　in　weight　δw　depends　on　the　updated　value　δx,y　increased　or　decreased　according　to　the　error,　in　order　to　reach　a　better　solution.　However,　the　previously　updated　values　are　neglected　after　every　iteration.　It　means　that　all　the　best　values　previously　achieved　in　weight　change　would　not　be　referring　back.　Hence,　there　is　no　information　sharing　between　the　best　values　that　have　been　achieved　at　the　previous　iterations　with　the　current　result.　Therefore,　by　using　the　term　＇global　best＇　concept　in　MRPROP,　the　information　of　previous　weight　change　is　the　only　guide　for　the　accurate　results.　Thus,　the　past　best　value　is　selected　in　term　of　optimized　solution　from　all　updated　values　of　the　current　weight　change　and　is　used　to　update　the　process.　This　variable　is　called　global　best　＇gbst＇.　The　gbst　selection　procedure　in　MRPROP　is:　First,　select　two　best　updated　values　randomly　from　all　the　current　change　in　weight　δw.　Then,　compare　these　two　values　in　term　of　optimized　solution　and　choose　the　better　one　as　gbst.　©　Springer　International　Publishing　AG　2017.