In　this　paper,　we　first　analyze　the　current　research　status　of　the　dimension　reduction　algorithm,　especially　the　subspace　learning.　We　find　that　the　global　subspace　learning　algorithm　is　based　on　the　assumption　that　the　data　is　distributed　in　the　same　space,　so　that　the　samples　contained　in　different　subspaces　are　not　well　represented.　In　order　to　overcome　this　shortcoming,　Researchers　have　proposed　the　methods　of　local　linear　subspace　learning　to　model　the　manifold.　But　no　matter　what　kind　of　algorithms　we　use,　we　must　determine　whether　the　samples　are　contained　in　different　subspaces.　If　samples　are　contained　in　different　subspaces,　the　number　of　subspaces　often　needs　to　be　given.　It　greatly　restricts　the　flexibility　and　performance　of　these　algorithms.　For　the　above　problems,　inspired　by　activation　function　in　neural　network,　an　adaptive　local　subspace　learning　method　is　proposed　in　this　paper.　Meanwhile　the　method　of　similarity　calculation　between　samples　is　given,　so　that　the　subspace　learning　of　the　sample　doesn＇t　need　to　consider　whether　samples　exists　in　a　number　of　different　subspaces.　Through　experiments　of　performance　analysis,　it　can　be　proved　that　the　subspace　learned　by　the　algorithm　can　be　better　at　representing　the　adjacency　relation　among　samples　in　manifold.　Experiments　of　recognition　and　reconstruction　on　face　database　verify　the　effectiveness　and　robustness　of　the　proposed　feature　extraction　method　on　high-dimensional　data　samples.