The multi-objective Neural Architecture Search (NAS) automates the process of neural network architecture design. It also evaluates and balances the accuracy and performance during the design process. The latency is used as a performance optimization object by many multi-objective NAS methods. A latency evaluation approach based on a latency prediction model can solve the efficiency problem of latency evaluation, but building a latency dataset of a certain scale for training the latency prediction model requires high computational costs. This paper uses multilayer perceptron (MLP) as the latency prediction model, takes an iterative selection of architectures to collect latency, trains the latency prediction model and proposes an active learning latency prediction model construction approach based on the performance of the latency prediction model to automatically stop collecting a large amount of latency data. It can avoid the high time cost problem. In the experiments, compared with the control group, using the active learning process to construct a latency prediction model for FPGA devices can save about 94% of the latency collection cost and achieve almost the same evaluation performance.