Learning beyond the synapse: activity-dependent myelination, neural correlations, and information transfer

Learning relies as much on synaptic weight modifications as it does on the timing of neural signaling. For instance, variability in axonal length and diameter may compromise spike-timing-dependent plasticity (STDP) by dispersing action potential timing, preventing synaptic potentiation and depression. Activity-dependent myelination (ADM), a form of brain plasticity relying on neuron-glia feedback, has been shown to normalize conduction times across different axons through adaptive, neuron-activity-dependent changes in axonal conduction velocity (CV). While historically seen as a mere compensation mechanism for the variable distances separating neurons, myelination is now recognized as a highly dynamic process supporting and imparting flexibility to various forms of neural computations ranging from learning, and memory consolidation to stimulus discrimination. We here employed a recurrent neural network of spiking Poisson neurons endowed with neuron-glia feedback and examined the functional consequences of ADM plasticity on population coding. We found that ADM plasticity implements a gain control mechanism up-regulating neuronal responses to dynamic stimuli, as well as tuning network correlations. Glia-mediated changes in CV along distinct axons were found to be titrated autonomously to promote AP coincidence, reinforcing the probability of post-synaptic response and hence to enhance information transfer. Such features were found to be preserved across spatial scales and to be axon specific, scaling with axon length, with longer axons supporting faster conduction compared with shorter ones. Neural integration relying on tight temporal precision, we exposed the network to dynamic inputs with different spatial and temporal statistics. ADM was found to enhance information transmission of such stimuli through axon-specific tuning of CV and the subsequent recruitment of stochastic resonance. Taken together, our results suggest that ADM represents an essential component of brain plasticity, coordinating neural responses to support the implementation of various forms of neural computations.