No document available.
Abstract :
[en] Neurons adapt their connections with each other through synaptic plasticity, driven by correlations in their spiking activity (Fig. 1A). Additionally, neuronal networks undergo global changes in rhythmic activity that correspond to different brain states, defined by switches in neuronal activity and orchestrated by neuromodulators. A well-known example is the transition from a state of learning during active waking to rest during quiet waking, which corresponds to a switch in neuronal activity from tonic firing to bursting (Fig. 1B). This raises the question of how switching from tonic firing to bursting affects the outcome of synaptic plasticity and whether it can support memory consolidation.
Recently, we have shown for a variety of synaptic plasticity models that bursting leads to a homeostatic reset, in which synaptic efficacy returns to a fixed baseline value irrespective of the starting point. This homeostatic reset causes the network to forget any learned information [1]. To address this issue, we propose an additional structural plasticity mechanism in which short-term changes in synaptic efficacy – evolving according to traditional plasticity rules – drive long-lasting morphological changes such as spine growth or insertion of new AMPA receptors. While synaptic efficacy undergoes homeostatic reset during bursting, information is consolidated through structural plasticity on a longer timescale.
We demonstrate the utility of this mechanism in a network of neurons using a conductance-based neuronal model that can switch from tonic firing to bursting along with a calcium-based synaptic rule to drive changes in synaptic efficacy. We investigate three regimes of switches in neuronal activity and plasticity mechanisms, denoted S1, S2, S3 (Fig. 1C). In S1, as a control condition, tonic firing is interleaved with periods of neuronal inactivity – mimicking bursting blockers – and a traditional plasticity rule, while in S2, tonic firing is separated by periods of bursting, leading to homeostatic reset in synaptic efficacy. Configuration S3 is identical as S2 but also includes our proposed burst-driven structural plasticity.
In our first memory task (Fig. 1D), we show that the signal-to-noise (SNR) is improved over repeated switches only in S3. In a simple pattern recognition task (Fig. 1E), blocking bursting activity (S1) makes the network fragile to noise, and blocking structural plasticity during bursting leads to complete forgetting (S2), neither of which occurs in S3. Finally, in a MNIST recognition task (Fig. 1F), we confirm that memory consolidation occurs with S3 by showing a stronger receptive field that consolidates during switches from tonic to burst and is robust to noise.
In this work, we shed light on the under-investigated role of switches in neuronal firing patterns for synaptic plasticity. Traditional plasticity rules result in a burst-induced homeostatic reset of synaptic efficacy, which is incompatible with memory consolidation. Our burst-driven structural plasticity proposes a solution to this problem, bridging the gap between switches in tonic firing to bursting, learning, and memory consolidation, and suggesting new ways to improve machine learning algorithms.
References
[1] Jacquerie K, Minne C, Ponnet J, et al. Switches to slow rhythmic neuronal activity lead to a plasticity-induced reset in synaptic weights. preprint, Biorxiv (2022).