Dynamics of systems with multiple stable or metastable states can be very sensitive to random perturbations. We show that by combining local randomly perturbed dynamical systems in a coupled network, one can preserve the attractors of the underlying local deterministic systems, while drastically reducing the effects of noise on the local dynamics. The mechanism of denoising is closely related to that of synchronization. We analyze both effects and discuss several applications to computational neuroscience.