The theory of cognitive diagnosis is deeply rooted in psychological and educational perspectives on human learning behavior. However, educators are faced with a challenging task: they must obtain effective reports from high-dimensional data inputs. Furthermore, they need to employ these reports to generalize their analyses to novel application scenarios. The publicly available methods to apply neural network to implement the parametric estimation are scare and often encounter technical difficulties in practice. We present LNN, a light-weight neural network to shape cognitive diagnostic in large-scale educational systems: blending theory with practicality. To validate the effectiveness of LNN, we conduct two types of experiments on the simulation datasets and real-world datasets. On both, LNN outperforms the existing approaches and drastically reduces the running costs, without suffering loss in accuracy and interpretability. The code for running LNN is open source and freely available.