Seol mar théacs é seo: Local minimization of prediction errors drives learning of invariant object representations in a generative network model of visual perception