부산대학교 도서관

Sequential behaviors such as language or bird songs are structured in time. This structure relies on the notion of long-distance-dependencies: transitions between words depend on the identity of words produced in the past. Here we propose a network mechanism supporting such dependencies. To do so we trained artificial neural networks to produce a minimal set of sequences exhibiting long-distance-dependencies. By reverse-engineering the trained networks we found this to rely on two superposing neural sequences, one responsible for the production of the motor sequence and another one encoding a contextual memory. We show how these two sequences are supported by neural activity and network connectivity and how they interact with each other to decide on transitions between words. We discuss similarities between the neural activity of our artificial neural networks and neural correlates of long-distance-dependencies that have recently been exposed in songbirds.