This paper develops a new neural network architecture for modeling spatial distributions (i.e., distributions on R^d) which is computationally efficient. The design of the architecture takes advantage of the specific structure of limit order books. The new architecture, which we refer to as a...