In this paper, we present a deep learning based method for blind hyperspectral unmixing in the form of a neural network autoencoder. We show that the linear mixture model implicitly puts certain architectural constraints on the network, and it effectively performs blind hyperspectral unmixing. Several different architectural configurations of both shallow and deep encoders are evaluated. Also, deep encoders are tested using different activation functions. Furthermore, we investigate the performance of the method using three different objective functions. The proposed method is compared to other benchmark methods using real data and previously established ground truths of several common data sets. Experiments show that the proposed method compares favorably to other commonly used hyperspectral unmixing methods and exhibits robustness to noise. This is especially true when using spectral angle distance as the network's objective function. Finally, results indicate that a deeper and a more sophisticated encoder does not necessarily give better results.