One　of　the　major　obstacles　in　using　deep　belief　network　(DBN)　is　the　structure　design.　Numerous　studies,　both　empirically　and　theoretically,　show　that　choosing　suitable　structure　can　improve　the　performance　of　DBN.　In　this　paper,　a　self-organizing　DBN　(S-DBN),　based　on　the　information　relevance　strategy　(IRS),　was　proposed　to　design　the　structure　of　DBN.　For　this　IRS,　the　maximal　information　coefficient　was　designed　to　measure　the　input　and　output　information　relevance　of　hidden　neurons.　Meanwhile,　the　mutual　information　was　introduced　to　measure　the　information　relevance　among　the　hidden　layers.　Then,　a　novel　self-organizing　strategy　was　developed　to　grow　and　prune　both　the　hidden　neurons　and　layers　during　the　training　process.　Moreover,　a　contrastive　divergence　algorithm　was　used　to　adjust　the　parameters　of　S-DBN.　Finally,　several　benchmark　problems　were　used　to　illustrate　the　effectiveness　of　S-DBN.　The　experimental　results　demonstrate　that　the　proposed　S-DBN　owns　better　performance　for　classification　problems　and　modeling　nonlinear　systems　than　some　existing　methods.　(C)　2019　Elsevier　B.V.　All　rights　reserved.