LargeConvolutionalNetworkmodelshaverecentlydemonstratedimpressiveclassicationperformanceontheImageNetbenchmark(Krizhevskyetal.,2012).Howeverthereisnoclearunderstandingofwhytheyperformsowell,orhowtheymightbeimproved.Inthispaperweaddressbothissues.Weintroduceanovelvisualizationtechnique