I am looking at non-linear dimensionality reduction techniques and am currently trying to understand the practical differences between different autoencoder approaches:
Can somebody point me to a comparison between RBM initialised autoencoders (Hinton) vs. de-noising autoencoders for dimensionality reduction?
Of course I am interesed in performace characteristics but also in more practical issues, such the reliability of finding a good solution and the computational cost of training.
Whilst not a direct answer to the comparison of RBM and Autoencoder, you might find the MATLAB toolbox for dimensionality reduction here very useful. The code is open source, there is an accompanying white paper, and autoencoders are one of the 34 techniques available.
The original stacked denoising autoencoder paper contains some comparisons to deep belief networks (i.e. stacked RBMs)
Vincent et al. (2010). Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion.
However, the comparisons are about classification performance, using the models in generative mode (to synthesize examples), and conceptual points (how the methods work). As far as I recall, they don't directly address your questions about computational cost, etc.
