On the role of non-linear latent features in bipartite generative neural networks