Structure from noise: Mental errors yield abstract representations of events

Humans are adept at uncovering complex associations in the world around them, yet the underlying mechanisms remain poorly understood. Intuitively, learning the higher-order structure of statistical relationships should involve sophisticated mental processes, expending valuable computational resources. Here we propose a competing perspective: that higher-order associations actually arise from natural errors in learning. Combining ideas from information theory and reinforcement learning, we derive a novel maximum entropy model of people's internal expectations about the transition structures underlying sequences of ordered events. Importantly, our model analytically accounts for previously unexplained network effects on human expectations and quantitatively describes human reaction times in probabilistic sequential motor tasks. Additionally, our model asserts that human expectations should depend critically on the different topological scales in a transition network, a prediction that we subsequently test and validate in a novel experiment. Generally, our results highlight the important role of mental errors in shaping abstract representations, and directly inspire new physically-motivated models of human behavior.