Memory consolidation facilitated by burst-driven late-phase plasticity

How do alternating periods of learning and rest contribute to memory consolidation? While it is recognized that learning relies on synaptic plasticity triggered by the spiking activity correlation between neurons, the role of rest periods and their biophysical mechanisms remain elusive. In this work, we leverage the interaction between the brain state fluctuations, reflecting changes in neuronal excitability, and memory, relying on synaptic plasticity occurring at different phases. Our approach involves a neural network model capable of transitioning between learning periods characterized by fast low-amplitude oscillations, and rest periods marked by slower large- amplitude oscillations. At the neuronal level, it is characterized by biophysical neurons capable of switching between input-driven tonic firing and the less-explored collective bursting.
In our model, synapses exhibit calcium-based early-phase plasticity, as studied in previous work. Here, we propose a new additional burst-induced late-phase plasticity mechanism. During learning, the early-phase plasticity forms new memories, as traditionally observed. During rest, the early-phase plasticity resets, returning to its baseline set point. It provides a physiological trace to drive the late-phase plasticity facilitating memory consolidation.
Validating our model through a memory task utilizing the MNIST dataset, we demonstrate that switching from tonic to burst, combined with early- and late-phase plasticity enables the network to acquire new information while preserving existing memories. The collective bursting activity during rest, combined with late-phase plasticity, represents the generation of new postsynaptic proteins and morphological synapse changes (termed structural plasticity). We find that substituting rest with an additional learning period impedes memory consolidation, rendering it susceptible to noise.
These findings propose a potential biological mechanism for unsupervised memory consolidation during rest and explain how the brain balances synaptic homeostasis and memory processes. Moreover, they suggest the utility of incorporating rest periods into machine learning models, highlighting the importance of including collective bursting and structural plasticity.