The success of deep learning in computer vision is partly attributed to the constructio nof large-scale labeled datasets such as ImageNet. However, conputer vision problems often require substantial human effrots and interventions to obtain accurate annotations due to dynamic aspects of class labels, needs for pixel-level labeling, and annotation ambiguities. Hence, colelcting high quality large-scale laveled datasets is very time consuming and even infeasible. This talk mainly addressess challenges and solutions related to data deficiency issue in the ocntext of semantic segmentation, where we can tak advantage of various types of weakly supervision. Specifically, I present a weakly supervised semantic segmentation algorithms with automatic data augmentation using videos. The proposed algorithm does not require any segmentation labels but show the state-of-the-art performance in the standard dataset. I also discuss weakly supervised learning in other problems including representation learning, multi-modal learning, online learning, etc.