Topic models have been prevalent for decades to discover latent topics and infer topic proportions of documents in an unsupervised fashion. They have been widely used in various applications like text analysis and context recommendation. Recently, the rise of neural networks has facilitated the emergence of a new research field -- Neural Topic Models (NTMs). Different from conventional topic models, NTMs directly optimize parameters without requiring model-specific derivations. This endows NTMs with better scalability and flexibility, resulting in significant research attention and plentiful new methods and applications. In this paper, we present a comprehensive survey on neural topic models concerning methods, applications, and challenges. Specifically, we systematically organize current NTM methods according to their network structures and introduce the NTMs for various scenarios like short texts and cross-lingual documents. We also discuss a wide range of popular applications built on NTMs. Finally, we highlight the challenges confronted by NTMs to inspire future research.