This talk is a part of the summarize of my researches which focus on the representation arrangement and manipulation in the unconditional and conditional perspectives. I will first conduct a brief introduce of my researches. After that, I will focus on a specific work I have done in unsupervised learning which is a instance of unconditional representation manipulation. The proposed paper is motivated by the absence of analysis for unsupervised image semantic segmentation (UISS). We build a connection between contrastive learning and UISS and design a paradigm for UISS. Experiments show that our model is effective and robustness.