Recovering　the　transmitted　signals　in　a　multiple-input　multiple-output　(MIMO)　system　is　known　to　be　non-deterministic　polynomial　hard.　It　is　extremely　challenging　to　obtain　a　high-quality　solution　with　fairly　low　computational　complexity.　To　solve　the　MIMO　detection　problem　effectively,　this　paper　proposes　to　model　it　as　a　time　series　prediction　problem,　and　a　bidirectional　temporal　convolutional　network　(Bi-　TCN)　is　designed　to　address　it.　In　Bi-　TCN,　the　encoder　extracts　the　features　of　the　received　signal　and　the　channel　matrix　by　applying　non-causal　dilated　convolution,　and　the　decoder　outputs　the　probability　distribution　of　the　recovered　transmitted　signal　in　parallel.　In　the　experiments,　we　compare　it　with　traditional　and　deep　learning-based　detectors　on　both　i.i.d.　and　correlated　Rayleigh　fading　channels,　respectively.　Experimental　results　empirically　demonstrate　that　Bi-　TCN　can　achieve　near-optimal　bit-error-rate　(BER)　performance　with　considerably　low　space　complexity.　©　2022　IEEE.