Dartmouth Events

Abstract: In the central nervous systems of many vertebrates, memories are reactivated and strengthened even in the absence of sensory input. Memory reactivation has been characterized as a mechanism for learning in evolutionarily complex nervous systems and has been explored in the context of singing in birds, navigation in rats, and visual perception in humans. Even in very simple nervous systems, however, there are distinct circuits involved in learning that may be susceptible to firing patterns that mimic reactivation. In this project, we use simplified neural networks to study computational properties of memory reactivation in two ways. First, we see how network topologies influence reactivation-like behavior. We find that a hierarchical network is more robust than a classical small-world network in generating reactivation-like behavior, noting that hierarchical networks may be more physiologically realistic for organisms. Second, we use simple machine learning methods to simulate a learning biological neural network and apply memory reactivation-based modifications to try to improve learning efficiency for a visual recognition task. We find that while reactivation-based modifications are not effective for this purpose, periodically downscaling weights globally in the network is effective. This conclusion is consistent with the synaptic homeostasis hypothesis, a current hypothesis that proposes a reason for sleep. Taken together, our findings may suggest that memory reactivation arose evolutionarily as a consequence of network topology in simpler organisms, and gained a role in learning after nervous systems diversified and became developmentally complex