It has been demonstrated that the neural population activity is often dominated by a few prominent latent neural modes in low-dimensional space that capture a significant fraction of the population covariance. Previous studies showed that those latent modes can be used as basic building blocks for brain-machine interfaces (BCI). In fact, it was demonstrated that through modeling those latent modes into variables, one can drive an online BCI with a performance surpassing traditional neural-observation-driven BCI. Additionally, it has been proposed that these neural modes are stable over long periods of time with respect to the intended behavior. To that end, we propose in this thesis methods and techniques that aim to enhance the performance of these latent-variable-driven BCIs in a multitude of ways. In the first chapter, we propose an unsupervised compression technique for neural interfaces that aims to compress input neural features by minimizing the loss of information between observations and modeled latent modes. Our results suggest the potential of design neural interfaces with sublinear power scaling with number of input electrodes which could enable power-efficient large-scale neural recordings. In the second chapter, we utilize latent modeling in a multimodal study featuring Electrocorticography (ECoG) and calcium imaging to show that you can predict the activity of modality using the other. We predict the latent modes of the calcium response using spectral features of ECoG recording then project decoded modes into the observation space to reconstruct single-cell activity of calcium response. In the last chapter, we extend this work in multimodal analysis to propose a generalized framework that unifies multiple modalities. One can leverage the unique strength of each modality by projecting then into a common latent space, which can be utilized to drive a brain-machine interface more accurately.