In　the　extreme　learning　machine　(ELM)　network,　sigmoid　activation　function　is　usually　chosen　for　additive　hidden　neurons.　Therefore,　this　paper　replaced　this　activation　function　with　a　smooth　approximation　called　softplus　function.　Because　of　being　closer　to　the　biological　activation　model　and　having　certain　sparseness,　softplus　activation　function　can　further　optimize　network　performance.　In　order　to　have　a　better　classification　performance,　the　optimization　model　of　ELM　by　the　improved　Fisher　discriminative　analysis　was　restricted,　and　animproved　ELM　algorithm　was　proposed.　Thus　the　output　weights　can　be　obtained　analytically　and　are　more　conducive　for　classification.　Finally,　the　experiments　on　handwritten　digit　database　and　face　database　prove　the　feasibility　and　superiority　of　the　improved　ELM　algorithm.　©,　2015,　Beijing　University　of　Technology.　All　right　reserved.