Dual stochasticity in the cortex as a biologically plausible learning with the most efficient coding

Abstract

Neurons and synapses in the cerebral cortex behave stochastically. The advantages of such stochastic properties have been proposed in several works, but the relationship and synergy between the stochasticities of neurons and synapses remain largely unexplored. Here, we show that these stochastic features can be inseparably integrated into a simple framework that provides a practical and biologically plausible learning algorithm that consistently accounts for various experimental results, including the most efficient power-law coding of the cortex. The derived algorithm overcomes many of the limitations of conventional learning algorithms of neural networks. As an experimentally testable prediction, we derived the slow retrograde modulation of the excitability of neurons from this algorithm. Because of the simplicity and flexibility of this algorithm, we anticipate that it will be useful in the development of neuromorphic devices and scalable AI chips, and that it will help bridge the gap between neuroscience and machine learning.