This　paper　presents　a　novel　modular　neural　network　called　brain-like　multi-hierarchical　modular　network　(BMNN).　Unlike　most　of　the　traditional　modular　neural　network,　the　BMNN　has　a　brain-like　multi-hierarchical　structure　and　uses　a　collaborative　learning　approach.　In　BMNN　learning　process,　each　input　sample　is　learned　by　multiple　sub-sub-modules　in　different　sub-modules　and　the　learning　result　of　BMNN　is　the　integration　of　the　multiple　sub-sub-modules　learning　results,　which　helps　to　improve　the　BMNN＇s　learning　accuracy　and　generalization　ability.　The　learning　algorithm　of　the　sub-sub-modules　is　an　algebraic　method　which　greatly　improves　the　BMNN＇s　learning　speed.　Applied　BMNN　to　mine　gas　concentration　forecasting　based　on　the　practical　production　data,　the　forecasting　results　compared　with　BP　neural　network　and　RBF　neural　network,　the　experiment　results　show　the　validity　of　the　proposed　forecasting　method　and　can　provide　the　scientific　decision　for　the　safety　in　coal　mine　production.　©　2014　IEEE.