Discrete Attractor Dynamics in the Frontal Cortex Nature 566, pages212–217(2019) Discrete attractor dynamics underlies persistent activity in the frontal cortex Hidehiko K. Inagaki, et.al. Janelia Research Campus, HHMI, Ashburn, VA, USA [paraphrase] Short-term memories link events separated in time, such as past sensation and future actions. Short-term memories are correlated with slow neural dynamics, including selective persistent activity, which can be maintained over seconds. In a delayed response task that requires short-term memory, neurons in the mouse anterior lateral motor cortex (ALM) show persistent activity that instructs future actions. To determine the principles that underlie this persistent activity, here we combined intracellular and extracellular electrophysiology with optogenetic perturbations and network modelling. We show that during the delay epoch, the activity of ALM neurons moved towards discrete end points that correspond to specific movement directions. These end points were robust to transient shifts in ALM activity caused by optogenetic perturbations. Perturbations occasionally switched the population dynamics to the other end point, followed by incorrect actions. Our results show that discrete attractor dynamics underlie short-term memory related to motor planning. Short-term memory is the ability of the brain to maintain information over seconds. Neurons in the frontal cortex and related brain regions show persistent, or slowly varying, changes in spike rate that correlate with the maintenance of short-term memories. This neural correlate has been extensively studied in delayed response tasks in non-human primates and, more recently, in rodents. In a typical task, an instruction informs the type of action to be performed and a go cue determines the timing of action, but only after a delay epoch, during which animals maintain a memory of the instruction and/or plan a movement. Persistent delay activity that predicts future movements is referred to as preparatory activity. The ALM is part of a multi-regional network that mediates motor planning. A large proportion of ALM neurons exhibit preparatory activity that predicts licking direction. The dynamics of ALM neural population can be analysed in activity space, in which each dimension corresponds to the activity of one neuron. ALM activity during the delay epoch is approximately two-dimensional. The coding direction (CD) vector discriminates trial types (that is, right or left lick), and activity projected along the CD contains almost all direction-selective activity. The second dimension corresponds to non-selective slow ramping activity. In a variety of behavioural tasks, preparatory activity ramps up to a movement. By contrast, standard discrete attractor models show stationary activity once the fixed points are reached. Ramping dynamics in discrete attractor networks can be obtained by the tuning of network parameters to generate slow drift to the fixed points or a non-selective ramping input that moves fixed points apart over time (external ramping model. The rate of recovery from perturbations is expected to be independent of ramping dynamics in the external ramping model. The membrane potential dynamics was inconsistent with cell-autonomous mechanisms as a primary mechanism for persistent activity. During the delay epoch, activity funnelled towards one of two discrete end points. After perturbations, activity trajectories recovered to reach one of the end points. These experiments provide direct evidence for multiple discrete attractors as a mechanism that underlies short-term memory. Our task design only has two behavioural choices. This probably explains two stable end points in ALM dynamics. The attractor model can accommodate a large range of end points. It is possible that each learned movement corresponds to a discrete attractor. Testing this hypothesis represents an important area for future investigation. Previous studies have reported neural activity and behaviour consistent with discrete attractors. For example, selective persistent activity in prefrontal cortex of primates is robust to sensory distractors, and remains discrete even with graded sensory stimuli. Our perturbation experiments show that this robustness and discreteness are properties of circuits that involve the frontal cortex. Together, discrete attractor dynamics subserve short-term memory in the frontal cortex in a wide-range of behaviours.