Across sensory systems, complex spatio-temporal patterns of neural activity can arise following the appearance (ON) and disappearance (OFF) of simple stimuli. These transient responses share remarkable similarities across sensory modalities and have been shown to be particularly informative about stimulus identity, suggesting they may form the basis for computations based on transient dynamics. However, the mechanisms generating transient OFF responses have so far not been fully elucidated.
Here we examine a network mechanism that generates transient responses through recurrent interactions. Focusing on linear recurrent networks, we determine the conditions for the existence of strong transients and examine their computational properties. We identify a general criterion that allows us to distinguish two sharply separated dynamical regimes depending on properties of the connectivity matrix: a regime in which all inputs lead to decaying transients, and a regime in which specific inputs elicit strong transients. Stimuli that elicit strong transient responses are mapped into a specific output state during the dynamics and can therefore be decoded from the two-dimensional neural trajectories they evoke.
Our theoretical framework suggests that transient OFF responses may be generated through a network mechanism. We explore this hypothesis by analysing OFF population responses recorded from the auditory cortex of mice passively listening to pure tones.