
Recurrent neural networks can learn an internal model of sensory experiences by encoding the statistical structure of stimulus-evoked activity patterns into their spontaneous activity patterns.


coremsg

The brain is thought to learn an internal model of the environment for improved performance in perception, decision making, and inference. Evidence suggests that spontaneous cortical activity represents such a model, or prior distribution, by cycling through stimulus-evoked activity patterns at frequencies proportional to the probabilities that these stimuli were previously experienced. However, how the brain encodes priors into spontaneous activity and utilizes them for inference tasks remains unclear. Here, we present a synaptic plasticity mechanism to generate cell assemblies encoding the statistical structure of salient sensory events and spontaneously replay these assemblies in spiking recurrent neural networks. The plasticity mechanism installs a Hebbian-like learning rule at excitatory and inhibitory synapses to minimize mismatches in the probability structure between stimulus-evoked and internally driven activities. Our model replicates the behavioral biases of monkeys performing perceptual decision making with surprising accuracy, demonstrating how spontaneous replay of previous experiences biases cognitive behaviors.

### Competing Interest Statement

The authors have declared no competing interest.

Learning rules for cortical-like spontaneous replay of an internal model

learning-rules-for-encoding-and-spontaneously-replaying-an-internal-model-of-sensory-experiences-in-cortical-networks


Learning Rules for Encoding and Spontaneously Replaying an Internal Model of Sensory Experiences in Cortical Networks
