With　the　arrival　of　the　big　data　era,　data　come　frequently　with　increasing　volume　and　high　dimensionality,　which　imposes　a　considerable　challenge　on　data　compression　in　network　representation　and　analysis.　How　to　learn　an　effective　low-dimensional　representation　has　a　dramatic　influence　on　performance　of　specific　network　learning　tasks.　In　this　paper,　we　propose　an　unsupervised　embedding　learning　feature　representation　scheme　by　deep　Siamese　neural　networks,　aiming　to　learn　an　efficient　low-dimensional　feature　subspace.　Unsupervised　embedding　learning　is　one　tough　but　interesting　task　in　which　its　searching　strategy　is　performed　without　the　guidance　of　class　label　information.　Siamese　network　is　a　neural　network　that　can　learn　an　efficient　feature　subspace　in　a　supervised　mode.　It　trains　two　networks　with　shared　weights　simultaneously,　and　feeds　them　with　random　sampling　from　the　same　dataset.　As　a　result,　the　feature　space　is　projected　onto　a　low-dimensional　subspace　such　that　the　similar　samples　are　with　small　values　close　to　zero,　whereas　the　dissimilar　ones　come　with　big　values　greater　than　a　predefined　margin.　We　further　discuss　deep　Siamese　neural　network　in　an　unsupervised　mode　and　its　applications　to　embedding　learning.　The　proposed　method　can　be　also　used　to　address　semi-supervised　feature　representation　problems.　Finally,　the　learned　unsupervised　embedding　is　validated　on　eight　publicly　available　databases　including　images,　voices,　and　text　documents.　Extensive　experiments　demonstrate　the　superiority　of　the　proposed　method　against　the　compared　existing　state-of-the-art　embedding　approaches.　©　2013　IEEE.