Bayesian　finite　element　model　updating　has　become　an　important　tool　for　structural　health　monitoring.　However,　it　takes　a　large　amount　of　computational　cost　to　update　the　finite　element　model　using　the　Bayesian　inference　methods.　The　surrogate　modeling　techniques　have　received　much　attention　in　recent　years　due　to　their　ability　to　speed　up　the　computation　of　Bayesian　inference.　This　study　introduces　two　new　surrogate　models　for　Bayesian　inference.　Specifically,　the　radial　basis　function　neural　networks　and　fully-connected　neural　networks　are　used　to　construct　surrogate　models　for　the　intractable　likelihood　function,　avoiding　the　enormous　computational　cost　of　repeatedly　calling　the　finite　element　model　in　the　Monte　Carlo　sampling　process.　A　full-scale　numerical　simulation　of　a　concrete　frame　and　a　six-story　steel　frame　experiment　were　selected　as　case　studies.　The　trained　surrogate　models　were　used　for　Bayesian　model　updating,　and　the　updated　results　were　compared　with　the　results　obtained　directly　using　the　finite　element　model　evaluation.　The　posterior　distributions　of　the　finite　element　model　parameters　obtained　using　the　trained　surrogate　models　are　sufficiently　accurate　compared　to　those　obtained　using　direct　finite　element　evaluation.　In　addition,　using　surrogate　models　for　finite　element　model　updating　greatly　reduces　computational　costs.