Cognitive communication and computing have seen deep penetration in many networking areas in the past decades. With the recent advances in big data analysis and deep learning, we have seen great potential toward exploring cognitive intelligence for a wide range of applications. A notable example therein is human activity recognition, especially through RFID. Existing RFID activity identification solutions are mostly designed for static or slowly moving targets, rendering them far from satisfactory. More importantly, we observe that they suffer serious performance degradation in typical indoor environments with multipath interference. In this article, we argue that the recent advance of deep learning brings new cognitive intelligence for human activity identification. We first review the literature and research challenges of multipath effects in indoor environments. Then we introduce an advanced RFID activity identification framework, DeepTag, which uses a deep-learning-based approach for activity identification in multipath-rich environments. DeepTag gathers massive phase information from multiple tags, and preprocesses them to extract such key features as pseudospectrum and periodogram. We feed the preprocessed signal power and angle information into a deep learning architecture that combines a convolutional neural network and long short-term memory (LSTM) network. Our DeepTag framework can well adapt to both tag-attached and tag-free activity identification scenarios. Our extensive experiments further demonstrate its superiority in activity identification in multipath-rich environments.