Device-free wireless localization and gesture recognition have received more and more attention in recent years, which can detect the location and gesture of a target by analyzing its shadowing effect on surrounding wireless links. Existing works have explored various statistical features extracted from received wireless signal to characterize the influence of the target on wireless links. However, to date, there is no universal method on feature selection and extraction for wireless localization and gesture recognition. Researchers often manually choose features in a subjective manner based on own experience, which is time-consuming and inaccurate. In this talk, a new deep learning based framework for device-free wireless localization and gesture recognition is presented. An auto-encoder network is constructed for feature extraction, which can automatically learn universal and discriminative features from received wireless signal without bias. A softmax regression based classifier is also developed to detect location and recognize gesture with the discriminative features learned from the auto-encoder network, which significantly increases the detection accuracy as compared with traditional methods using handcraft features.