Embedded chimera states in recurrent neural networks

Abstract

Fully and partially synchronized brain activity plays a key role in normal cognition and in some neurological disorders, such as epilepsy. However, the mechanism by which synchrony and asynchrony co-exist in a population of neurons remains elusive. Chimera states, where synchrony and asynchrony coexist, have been documented only for precisely specified connectivity and network topologies. Here, we demonstrate how chimeras can emerge in recurrent neural networks by training the networks to display chimeras with machine learning. These solutions, which we refer to as embedded chimeras, are generically produced by recurrent neural networks with connectivity matrices only slightly perturbed from random networks. We also demonstrate that learning is robust to different biological constraints, such as the excitatory/inhibitory classification of neurons (Dale’s law), and the sparsity of connections in neural circuits. The recurrent neural networks can also be trained to switch chimera solutions: an input pulse can trigger the neural network to switch the synchronized and the unsynchronized groups of the embedded chimera, reminiscent of uni-hemispheric sleep in a variety of animals. Our results imply that the emergence of chimeras is quite generic at the meso- and macroscale suggesting their general relevance in neuroscience.