Cuckoo search algorithm is a simple and efficient meta-heuristic algorithm, while it can be easily trapped into local optimum when solving complex multimodal optimization problems. To tackle this problem, an improved cuckoo search algorithm based on neural networks was proposed by combining the characteristics of neural network algorithm and cuckoo search algorithm. The core idea of this algorithm is to balance global search ability and local search ability of cuckoo search algorithm with powerful global search ability of the improved neural network algorithm and dynamic population strategy, thereby reducing the possibility of the cuckoo algorithm falling into local optimum. The algorithm firstly sorts the individuals in the population according to the fitness values. Then the best half individuals of the population are performed by the cuckoo search algorithm, whereas the worst half individuals are optimized by the improved neural network algorithm. Finally all individuals are grouped into a new population, from which the optimal solution can be selected. In this experiment, 24 complex multimodal optimization problems were employed to study the optimization performance and compare between the proposed algorithm and neural network algorithm, cuckoo search algorithm, and other improved cuckoo algorithms. Results show that the proposed algorithm fully demonstrated the advantages of the modified neural network algorithm and the cuckoo search algorithm, which was significantly better than other algorithms in resolution quality, convergence speed, and stability.