Deep　belief　network　applying　unsupervised　methods　of　greedy　layer　training,　from　the　training　set　automatic　feature　extraction　value,　will　cause　the　error　by　layer　transfer,　thus　affecting　the　accuracy　of　the　model　prediction,　in　order　to　solve　this　problem,　proposed　using　conjugate　gradient　algorithm　in　gradient　descent　can　accelerate　the　convergence　of　ideas,　improvement　of　the　restricted　Boltzmann　machine　network　algorithm　in　the　depth　of　confidence,　first　from　the　complexity　of　the　algorithm　and　the　reconstruction　error　analysis　of　improved　model　differences　and　advantages,　and　classify　the　verification　on　the　MNIST　data　set,　and　a　detailed　analysis　of　the　feasibility　of　the　improved　model　and　efficiency,　the　experimental　results　shows　the　feature　extraction　ability　improved　deep　belief　network　model　has　better　and　classification　results.　©　2018　IEEE.