In extracting statistical regularities from the seemingly random environment, our minds grow special interests in patterns. To account for such a behavior, much research has been focusing on top-down influences such as the representativeness heuristic and Bayesian belief updating. Here we take a reverse-engineering approach by first examining the waiting time statistics and the self-overlap property of patterns and revealing a normative basis for people's special attention to patterns. With a unsupervised neural network simulation, we show that different patterns may leave different traces in mind corresponding to the waiting time statistics, indicating an early pattern dissociation without any top-down guidance. We argue that the sense of randomness could have started locally with short sequences and emerged early at the perceptual level, and, the process of spatial-temporal association may be the early driving force towards a structured hypothesis space.