There have been more and more Internet-of-Things (IoT) applications emerged in the recent few years, e.g., connected robots, Internet-of-Vehicles, IoT-enabled Smart Grid, IoT-based Sensing, etc. These new-generation IoT applications are potential to totally revolutionize people’s everyday life. However, before that, a series of new artificial intelligence (AI) methods are needed to efficiently analyze the big, heterogeneous IoT data, and automatically give reliable decisions, accurate predictions, or quick yet correct feedbacks. In the dissertation, the author focused on the design and implementation of some novel deep learning approaches, in order to address various challenging problems in several emerging IoT applications. First, the author proposes a view-invariant Convolutional Neural Network (CNN) Model for the scene understanding tasks of connected disaster-handling robots. In this system, two individual CNNs are used to, respectively, propose objects from input data and classify their categories. The author attempts to overcome the difficulties and restrictions caused by disasters using several specially-designed multi-task loss functions. The most significant advantage in this work is that the proposed method can learn a view-invariant feature with no requirement on RGB data, which is essential for harsh, disordered and changeable environments. Second, the author adopts AI methods to implement intelligent decision-making for autonomous vehicles. A human-like driving system is proposed to give autonomous vehicles the ability to make decisions like a human. In this method, a decision-making system calculates the specific commands to control the vehicles based on the abstractions. The biggest advantage of this work is that the author implements a decision-making system which can well adapt to real-life road conditions, in which a massive number of human drivers exist. Third, the author focuses on the electrical load forecasting task for the Internet-of-Energy. An IoT-based deep learning system is introduced to automatically extract features from the captured data, and ultimately, give an accurate estimation of future load value. One significant advantage of this method is the specially designed two-step forecasting scheme, which significantly improves the forecasting precision. Also, the proposed method is able to quantitatively analyze the influences of some major factors.