Among other aspects, a storage medium is provided that stores executable instructions by a processor to implement a method. The method includes generating feature representations including distributions on a set of features corresponding to objects of an object training corpus. The method further includes deriving a subject model defining a set of subjects by performing (30) a latent Dirichlet allocation (LDA) with a prior probability distribution of Dirichlet composed by an Indian buffet process (IBP), the deduction being performed using a reduced Gibbs sampling algorithm by iteratively sampling (1) LDA subject allocation variables and (2) binary activation variables from the prior Dirichlet probability distribution composed by an IBP.
展开▼