Short term memory by transient oscillatory dynamics in recurrent neural networks

Despite the importance of short-term memory in cognitive function, how the input information is encoded and sustained in neural activity dynamics remains elusive. Here, by training recurrent neural networks to short-term memory tasks and analyzing the dynamics, the characteristic of the short-term memory mechanism was obtained in which the input information was encoded in the amplitude of transient oscillation, rather than the stationary neural activities. This transient orbit was attracted to a slow manifold, which allowed for the discarding of irrelevant information. Strong contraction to the manifold results in the noise robustness of the transient orbit, accordingly to the memory. The generality of the result and its relevance to neural information processing were discussed.