This paper discusses the challenges of using unsupervised learning to capture significant factors of variation and similarity, as there is no clear cost function that can do so. However, since natural systems tend to exhibit smooth dynamics, failing to incorporate this feature into the unsupervised objective function represents a missed opportunity. To better leverage the gradual and uncertain knowledge gained through pseudo-supervision, this paper proposes the use of dynamic objective functions, which are more effective than classical static cost functions. Specifically, this paper introduces Improved Dynamic Autoencoder (IDynAE), a novel model for deep clustering that computes the conflicted and unconflicted sample based on the topological filter. When we used the topological filter, the topology information of the cluster model is preserved, which leads to more accurate estimation of the conflicted and unconflicted samples in order to estimate the with results of 84.36% and 40.97% in terms of ACC and NMI respectively.