I've read two interesting research papers on face recognition and I'd like to understand the relation between them.
- "Learning to Disentangle Factors of Variation with Manifold Interaction" by Reed et al.
- "Tensor Analyzers" by Tang et al.
In Reed et al. they say "Our approach is complementary [...], and is also capable of exploiting correspondence information". By correspondence information, they mean using related data points. For example, two different pictures of the same person's face correspond to each other by identity.
How are the approaches complementary?
Tang et. al seem to be using a purely statistical method which can be thought of as an extensions to Factor Analysis. Basically, instead of having a single latent variable, they have groups of latent variables with gaussian priors. They use the resulting Tensor Loading for classication.
Reed et. al use Restricted Boltzman Machines (RBMs) but put restrictions on the energy function of their latent variables that ensure disentangling.
Regardless, both approaches seem to be trying to find patterns accross corresponding information, but how are they complementary? What evidence is there that they aren't identifying the same things?