|
While trying to zero in on a suitable architecture for the domain concerned, how can we evaluate the pros and cons for each architecture? I am particularly looking at unsupervised feature learning methods. I tried to list down the broad categories: RBM, Sparse Coding, GMM -> DBN, DBM; And other variants like convolutional, gated etc. AE -> SAE; And many variants of these like Denoising/Sparse/Convolutional etc. I went through the Review paper on Representation Learning, but I need a little more help on understanding the key differences in the above mentioned models, and how to evaluate their relevance to a particular domain? [Please feel free to add other variants of these models, which I missed in the question] I have also seen at a few places that there is not much difference between them, except for the additional exponentiation in the probabilitic models when compared to neural networks. Can anyone either give a detailed summary or a pointer towards one, where key characteristics of each of these architectures and their applications are described? Apologies if I am asking for a lot :D! |
