The main task of granular computing (GrC) is about representing, constructing, and processing information granules. Information granules are formalized in many different approaches. Different formal approaches emphasize the same fundamental facet in different ways. In this paper, we propose a novel GrC method of machine learning by using formal concept description of information granules. Based on information granules, the model and mechanism of two-way learning system is constructed in fuzzy datasets. It is addressed about how to train arbitrary fuzzy information granules to become necessary, sufficient, and necessary and sufficient fuzzy information granules. Moreover, an algorithm of the presented approach is established, and the complexity of the algorithm is analyzed carefully. Finally, to interpret and help understand the theories and algorithm, a real-life case study is considered and experimental evaluation is performed by five datasets from the University of California-Irvine, which is valuable for applying these theories to deal with practical issues.