That depends on the nature of the data, I think. If the data has a lot of sequential, sparse, hierarchical statistical dependencies (like source code, text or d

That depends on the nature of the data, I think. If the data has a lot of sequential, sparse, hierarchical statistical dependencies (like source code, text or data streams), they might be better modeled by an LSTM. If you have high-dimensional dependencies (like images, where each pixel tends to spatially depends on many other pixels), then an autoencoder or some undirected model might be the right choice.