you might want to take a look at http://www.stat.cornell.edu/~li/abc_empirical.pdf but it is about tree boosting algorithms (evaluated on UCI "Covertype", "Letter", "Poker") and comparison with deep learning algorithms is taken from the results of http://www.iro.umontreal.ca/~lisa/seminaires/13-06-2007.pdf (on MNIST "Digits") ...
This is not exactly what your asking for, but check out
Mitchell, Ben, and John Sheppard. "Deep Structure Learning: Beyond Connectionist Approaches." Machine Learning and Applications (ICMLA), 2012 11th International Conference on. Vol. 1. IEEE, 2012.
They demonstrate "the impact of the deep architecture itself" using only linear methods.