Switches in neuronal activities: what are the impact on memory?

Neurons adapt their connections with each other through synaptic plasticity, driven by correlations in their spiking activity. Furthermore, neuronal networks may undergo global changes in rhythmic activity, corresponding to different brain states defined by switches in neuronal excitability, orchestrated by neuromodulators. A well-known example is the transition from a state of learning to quiet wakefulness or sleep, corresponding to a neuronal activity switch from tonic firing to bursting, known to support memory consolidation [Dastgheib et al, 2022].
Recently, we have shown for a variety of synaptic plasticity models that bursting leads to a homeostatic reset, in which synaptic weights return to a fixed baseline value irrespective of the starting point, therefore forgetting any learned information [Jacquerie et al,2022] . To address this issue, we propose an additional structural plasticity mechanism [Poirazi and Mel, 2001] in which short-term changes in synaptic efficacy – evolving according to traditional plasticity rules – drive long-lasting morphological changes such as spine growth or receptor density upregulation (AMPAfication). While synaptic efficacy undergoes homeostatic reset during bursting as before, information is consolidated through structural plasticity. It can then be read out from an effective synaptic weight as the product of the efficacy and receptor density.
We demonstrate the utility of this mechanism in a network of neurons performing three memory tasks: (i) an associative memory task as done in [González-Ruedas et al, 2018], (ii) a pattern recognition task and (iii) a learning task on the MNIST dataset. For each task, information is encoded by applying external input currents during a tonic phase and consolidated during a subsequent bursting phase. Then, we computed either the signal-to-noise ratio or the learning performance throughout tonic firing and bursting cycles. We compared this performance when the network is burst-deprived. For each task, AMPAfication paired with burst-induced homeostatic reset improves signal-to-noise (SNR) over repeated learning trials.
Overall, this mechanism reconciles traditional plasticity rules with burst-induced homeostatic reset, makes important experimental predictions about the role of switching brain states, and even has the potential for novel machine-learning algorithms.