The next generation of mobile networks is set to become increasingly complex, as these struggle to accommodate tremendous data traffic demands generated by ever-more connected devices that have diverse performance requirements in terms of throughput, latency, and reliability. This makes monitoring and managing the multitude of network elements intractable with existing tools and impractical for traditional machine learning algorithms that rely on hand-crafted feature engineering. In this context, embedding machine intelligence into mobile networks becomes necessary, as this enables systematic mining of valuable information from mobile big data and automatically uncovering correlations that would otherwise have been too difficult to extract by human experts. In particular, deep learning based solutions can automatically extract features from raw data, without human expertise. The performance of artificial intelligence (AI) has achieved in other domains draws unprecedented interest from both academia and industry in employing deep learning approaches to address technical challenges in mobile networks. This thesis attacks important problems in the mobile networking area from various perspectives by harnessing recent advances in deep neural networks. As a preamble, we bridge the gap between deep learning and mobile networking by presenting a survey on the crossovers between the two areas. Secondly, we design dedicated deep learning architectures to forecast mobile traffic consumption at city scale. In particular, we tailor our deep neural network models to different mobile traffic data structures (i.e. data originating from urban grids and geospatial point-cloud antenna deployments) to deliver precise prediction. Next, we propose a mobile traffic super resolution (MTSR) technique to achieve coarse-to-fine grain transformations on mobile traffic measurements using generative adversarial network architectures. This can provide insightful knowledge to mobile operators about mobile traffic distribution, while effectively reducing the data post-processing overhead. Subsequently, the mobile traffic decomposition (MTD) technique is proposed to break the aggregated mobile traffic measurements into service-level time series, by using a deep learning based framework. With MTD, mobile operators can perform more efficient resource allocation for network slicing (i.e, the logical partitioning of physical infrastructure) and alleviate the privacy concerns that come with the extensive use of deep packet inspection. Finally, we study the robustness of network specific deep anomaly detectors with a realistic black-box threat model and propose reliable solutions for defending against attacks that seek to subvert existing network deep learning based intrusion detection systems (NIDS). Lastly, based on the results obtained, we identify important research directions that are worth pursuing in the future, including (i) serving deep learning with massive high-quality data (ii) deep learning for spatio-temporal mobile data mining (iii) deep learning for geometric mobile data mining (iv) deep unsupervised learning in mobile networks, and (v) deep reinforcement learning for mobile network control. Overall, this thesis demonstrates that deep learning can underpin powerful tools that address data-driven problems in the mobile networking domain. With such intelligence, future mobile networks can be monitored and managed more effectively and thus higher user quality of experience can be guaranteed.