Dual stochasticity of neurons and synapses for sampling-based learning in the brain

Neurons and synapses behave highly stochastically in the brain. However, how this stochasticity is beneficial for computation and learning in the brain remains largely unknown. In this presentation, we will see that the stochastic processes in neurons and synapses can be integrated into a unified framework to optimally sample events from the environments, resulting in an efficient learning algorithm consistent with various experimental results. In particular, the learning algorithm enables us to reproduce the recently discovered efficient power-law coding in the cortex. These results suggest that synapses and neurons work cooperatively to implement a fundamental method for stochastic computing in the brain.