Stimulus representation in mouse visual cortex neurons is highly plastic throughout perceptual learning. Task-relevant visual patterns become better represented by the neuronal population as behavioural performance improves. Existing theories of perceptual learning describe plasticity of stimulus representation as the contextual top-down modulation of bottom-up signals. Local synaptic connections account for a significant portion of excitatory current in layer 2/3 pyramidal cells in V1, while the number of these connections vary widely across neurons in the network. Neurons with a large number of recurrent synaptic connections are highly coupled to the population-wide activity, while relatively unconnected neurons have low population coupling. However, how local recurrent connectivity impact top-down contextual modulations remains unknown.

Using a computational recurrent network model of perceptual learning we study: 1) how does a neurons’ population coupling impact stimulus representation plasticity during perceptual learning, and 2) how can diverse population coupling emerge and be maintained throughout synaptic plasticity? We find that the distribution of neurons’ population coupling within the network has a dramatic effect on the network’s ability to form new stimulus associations from combining bottom-up and top-down information. Diverse population coupling ensures that there are both weakly coupled neurons with stable stimulus representations that faithfully reflect their bottom-up inputs, and strongly coupled neurons with stimulus representations that can be flexibly modulated by top-down signals during perceptual learning. Surprisingly, we find that diverse population coupling emerges in networks of neurons with diverse learning rates: neurons with slow learning rates maintain low population coupling and robust stimulus representations, whereas neurons with fast learning rates maintain a high population coupling, enabling them to flexibly learn new stimulus representations.