The paper introduces the Masked Autoencoder Distribution Estimator (MADE), a modification of autoencoder neural networks that enables them to estimate distributions from a set of examples. By masking the autoencoder's parameters to respect autoregressive constraints, each input is reconstructed only from previous inputs in a given ordering. This allows the autoencoder's outputs to be interpreted as conditional probabilities, and their product as the full joint probability. The method can be applied to various architectures, including deep ones, and is implemented efficiently on GPUs. Experiments demonstrate that MADE is competitive with state-of-the-art tractable distribution estimators, offering significant speed improvements and better scaling than other autoregressive estimators. The paper also explores extensions such as training MADE with multiple orderings and hidden layer connectivity structures, and evaluates these on binary datasets with hundreds of dimensions.The paper introduces the Masked Autoencoder Distribution Estimator (MADE), a modification of autoencoder neural networks that enables them to estimate distributions from a set of examples. By masking the autoencoder's parameters to respect autoregressive constraints, each input is reconstructed only from previous inputs in a given ordering. This allows the autoencoder's outputs to be interpreted as conditional probabilities, and their product as the full joint probability. The method can be applied to various architectures, including deep ones, and is implemented efficiently on GPUs. Experiments demonstrate that MADE is competitive with state-of-the-art tractable distribution estimators, offering significant speed improvements and better scaling than other autoregressive estimators. The paper also explores extensions such as training MADE with multiple orderings and hidden layer connectivity structures, and evaluates these on binary datasets with hundreds of dimensions.