Modeling Latent Attention Within Neural Networks