Stabilization of memory on neural manifolds through multiple synaptic time scales

Working memory is commonly thought to be implemented in the brain by the persistent activity of attractor neural networks. Compelling evidence exists for the implementation of such memory networks in vertebrate and invertebrate brains. Naively, retention of continuous memories in attractor networks – as observed in the head direction systems of mammals and insects and the oculomotor integrator in vertebrates – is particularly sensitive to neural stochasticity, because noise can continuously shift the activity of the network along its manifold of steady states. As these random drifts accumulate over time, the stored memory degrades. Small networks, such as the head direction system of insects, are particularly prone to this form of memory deterioration, and it is unclear how such networks can maintain their persistent state over behaviorally relevant timescales. Here we identify a neural mechanism that effectively counteracts random drifts by employing a combination of slow excitatory and fast inhibitory synaptic connections. We show that the proposed mechanism can substantially decrease diffusivity in linear and nonlinear networks. Hence, the mechanism offers a plausible explanation for how the brain can stably store memories of continuous quantities, despite the ubiquity of noise. Finally, we identify how to engineer connectivity such that variability in neural activity perpendicular to the attractor is unaffected by the stabilization mechanism. This scheme allows for tight confinement of neural activity patterns to a low dimensional manifold, as well as the rapid relaxation of transient modes of activity back into the attractor. The theory suggests that neurons in head direction cell networks that are commonly thought to be utilized for velocity integration may also aid in stabilization against noise-driven motion.