Extreme learning machines (ELMs) have been proposed for generalized single-hidden-layer feedforward networks which need not be neuron alike and perform well in both regression and classification applications. The problem of determining the suitable network architectures is recognized to be crucial in the successful application of ELMs. This paper first proposes a dynamic ELM (D-ELM) where the hidden nodes can be recruited or deleted dynamically according to their significance to network performance, so that not only the parameters can be adjusted but also the architecture can be self-adapted simultaneously. Then, this paper proves in theory that such D-ELM using Lebesgue p-integrable hidden activation functions can approximate any Lebesgue p-integrable function on a compact input set. Simulation results obtained over various test problems demonstrate and verify that the proposed D-ELM does a good job reducing the network size while preserving good generalization performance.