We study the cooperative control problem for a group of autonomous surface vehicles (ASV) with uncertain dynamics. A new decentralized cooperative controller is developed for a group of underactuated surface vehicles by employing the neural network-based dynamic surface approach, graph theory and Lyapunov stability theory. Using this design, it does not require to calculate the numerical derivatives of the virtual control signals as in traditional backstepping-based design. The advantages of the proposed cooperative controller are that, in addition to achieve a desired formation, the unknown uncertain dynamics such as coriolis and centripetal force, hydrodynamic damping, unmodelled hydrodynamics, disturbances from environment can be compensated by on-line learning. An illustrative example is provided to demonstrate the effectiveness of the proposed approach.