Despite rapidly growing interest in harnessing machine learning in the study of quantum manybody systems, training neural networks to identify quantum phases is a nontrivial challenge. The key challenge is in efficiently extracting essential information from the many-body Hamiltonian or wave function and turning the information into an image that can be fed into a neural network. When targeting topological phases, this task becomes particularly challenging as topological phases are defined in terms of non-local properties. Here we introduce quantum loop topography (QLT): a procedure of constructing a multi-dimensional image from the "sample" Hamiltonian or wave function by evaluating two-point operators that form loops at independent Monte Carlo steps. The loop configuration is guided by characteristic response for defining the phase, which is Hall conductivity for the cases at hand. Feeding QLT to a fully-connected neural network with a single hidden layer, we demonstrate that the architecture can be effectively trained to distinguish Chern insulator and fractional Chern insulator from trivial insulators with high fidelity. In addition to establishing the first case of obtaining a phase diagram with topological quantum phase transition with machine learning, the perspective of bridging traditional condensed matter theory with machine learning will be broadly valuable.Introduction-Machine learning techniques have been enabling neural networks to successfully recognize and interpret big data sets of images and speeches [1]. Through supervised trainings with a large number of data sets, neural networks 'learn' to recognize key features of a universal class. Very recently, rapid and promising development has been made from this perspective on numerical studies of condensed matter systems, including dynamical systems[2-6], systems undergoing phase transitions [7][8][9][10][11][12][13], as well as quantum many-body systems. Also established is the theory connection to renormalization group [14,15]. Exciting successes in application of machine learning to symmetry broken phases [7][8][9][10] may be attributed to the locality of the defining property of the target phases: the order parameter field. The snap-shots of order parameter configuration form images that can be readily fed into neural networks that have been developed to recognize patterns in images.Unfortunately many novel states cannot be numerically detected through a local order parameter. For one, all topological phases are intrinsically defined in terms of non-local topological properties. Not only many-body localized states of growing interest [16] fit into this category, even a superconducting state fits in here since the superconducting order parameter explicitly breaks particle number conservation [17]. In order for neural networks to learn to recognize and identify such phases, we need to supply them with "images" that contain relevant non-local information. Clearly information based on single site is insufficient. One approach to detecting topologic...