Brain-Computer Interface (BCI) bridges the human being's neural world and the outer physical world by decoding individuals' brain signals into commands recognizable by computer devices, which has attracted increasing attention in recent years. This dissertation aims at overcoming the hurdles and stretching the horizons of data-efficient BCI systems by developing robust deep representation learning paradigms. First, we present a comprehensive introduction of BCI systems including the deep learning models, state-of-the-art studies adopting deep learning for BCI drawbacks, and the appealing real-world BCI applications. Moreover, we propose automatic high-level representation learning methods through deep architectures addressing the traditional time-consuming manually feature engineering and low signal-to-noise ratio data. In addition, we develop reinforced selective attention mechanism by combining reinforcement learning and deep neural network for capturing informative representations adaptive to different scenarios. Furthermore, we design a weakly-supervised predictive model to harness the deep generative model and generative adversarial networks collectively under a trainable unified framework addressing the shortage of labeled data. At last, upon the proposed models, we develop several real-world BCI applications such as an EEG-based person identification system and a prototype of a brain-controlled typing system which converts user's thoughts into text.