Landslide risk assessment (LRA) is of great significance to reduce the loss of landslide disaster and improve the effect of disaster prevention and reduction. At present, there are few studies on LRA, and the theories and methods need to be innovated. In view of this, convolutional neural network (CNN) is applied to LRA. Based on 9 landslide hazard conditioning factors and 5 vulnerability conditioning factors, a CNN framework suitable for LRA (LR-CNN framework) was constructed by using Bayesian method to optimize the CNN hyperparameters. In order to obtain more stable model and assessment results, 5-fold cross-validation method was adopted. The LR-CNN framework and the risk expression of the United Nations Department of human affairs (risk expression of UNDHA) method were applied to Tumen City, China, and the two methods were analyzed and compared. From the rationality test results, both methods meet the rationality test standard. From the landslide risk mapping, both methods can obtain better LRA results. From the area within each risk class, LR-CNN framework is slightly inferior to risk expression of UNDHA method, but from the perspective of landslide distribution and landslide density, LR-CNN framework is superior to risk expression of UNDHA method. It can be seen that the application of CNN to LRA is feasible and superior to the risk expression of UNDHA method to a certain extent.