The uncanny valley phenomenon refers to the deep sense of unease caused by human-like robots, often attributed to the "difficulty of categorization" hypothesis. This study proposes a Bayesian model, grounded in the brain's prediction error minimization principle, to explain this hypothesis by formulating affinity in terms of negative entropy. Additionally, we analyzed the impact of prediction and observation uncertainty on affinity. Through a subject experiment, we demonstrated that blurring a robot's image—thereby increasing observation uncertainty—can mitigate the decline in affinity. This study introduces a mathematical model of the uncanny valley and highlights prediction and observation uncertainty as novel design variables in robot development.