Deep Belief Networks vs Convolutional Neural Networks performance on non-Image Classification Tasks

Question

In the paper Improved Classification based on Deep Belief Networks, the authors have stated that for better classification, generative models are used to initialize the model and model features before training a classifier. Typically they are needed to solve separate unsupervised and supervised learning problems. Generative restricted Boltzmann machines and deep belief networks are widely used for unsupervised learning purposes.

My question is that, if I was to perform a non-image multi-class classification task through unsupervised learning, would it be better to use Deep Belief Networks or Convolutional Neural Networks without considering the fact that dataset matters as well?

A similar question related to image-classification tasks was asked here Deep Belief Networks vs Convolutional Neural Networks. The answer stated that DBNs are likely to perform better for non-image classification tasks than CNNs, but is there any evidence available regarding this, or any paper that explores this more deeply?

Answer 1

The operations in a convolutional neural network are specifically tuned towards image processing. E.g the feature extraction convolution with parameter sharing is run on different parts of the image, also CNNs include subsampling layers which can be understood as producing smaller versions of the (processed) input image. Because of this, I would imagine that CNNs have an inherent disadvantage if the input data is not an image or sufficiently image-like.