Sensory events generally exhibit a certain degree of uncertainty in the real world. For optimizing behavioral outputs, the brain has to remember these events together with the probabilities of their occurrences and utilize the knowledge when necessary. Although spontaneous activity is thought to play an active role in this probabilistic computation, the underlying neural mechanisms have been poorly understood. In this talk, I will propose a mechanism in which a recurrent neural network segregates and remembers the multiple sensory events that repeat in temporal input with different probabilities. We first describe a self-supervised learning rule that enables a two-compartment neuron to segregate repeated patterns in temporal synaptic input. Then, we demonstrate that similar learning rules implemented at excitatory and inhibitory synapses allow a recurrent neural network to spontaneously reactivate the evoked response patterns with the frequencies proportional to the experienced probabilities of the evoked responses. We ask our network model to perform a sensory-guided alternative choice task and show that the model can account for the biases in choice responses observed previously in monkeys surprisingly well. Our results describe computational principles in which the brain spontaneously replays experiences with accurate probabilities and suggest how such replays bias our perception of sensory events to generate human-like behaviors. This work was supported by KAKENHI no. 18H05213 from JSPS, and most results were obtained in collaboration with Toshitake Asabuki.