A Unified Dynamic and Computational Mechanism for Persistent and Transient Neural Activity Patterns During Delayed-Response Tasks

Neural activity during short-term memory can be either persistent or transient. However, there is an ongoing debate about which activity pattern presents a more accurate neural basis underlying short-term memory. Here, we addressed this problem by training artificial recurrent neural networks (RNNs) with delayed-response tasks. We found biological features emerged from the trained RNNs. Reverse-engineering showed that persistent and transient neural activity patterns can be unified by a neural state moving on a low-dimensional heteroclinic manifold, driven by a velocity field which is computationally Bayes-optimal to minimize the memory error. Our results shed new light on the neural mechanism and computational principles of short-term memory by unifying two seemly contradictory experimental phenomena into a single framework and suggest a new way to interpret the experimental data using computationally important transient speed as a continuous measure to characterize the neural activities during the delay.