Embedding stochastic dynamics of the environment in spontaneous activity by prediction-based plasticity

Toshitake Asabuki, Claudia Clopath

Abstract

The brain learns an internal model of the environment through sensory experiences, which is essential for high-level cognitive processes. Recent studies show that spontaneous activity reflects such a learned internal model. Although computational studies have proposed that Hebbian plasticity can learn the switching dynamics of replayed activities, it is still challenging to learn dynamic spontaneous activity that obeys the statistical properties of sensory experience. Here, we propose a pair of biologically plausible plasticity rules for excitatory and inhibitory synapses in a recurrent spiking neural network model to embed stochastic dynamics in spontaneous activity. The proposed synaptic plasticity rule for excitatory synapses seeks to minimize the discrepancy between stimulus-evoked and internally predicted activity, while inhibitory plasticity maintains the excitatory-inhibitory balance. We show that the spontaneous reactivation of cell assemblies follows the transition statistics of the model's evoked dynamics. We also demonstrate that simulations of our model can replicate recent experimental results of spontaneous activity in songbirds, suggesting that the proposed plasticity rule might underlie the mechanism by which animals learn internal models of the environment.

Introduction

The brain does not merely record sensory input; it constructs an internal model of the world to predict future events. This internal model is often manifested in "spontaneous activity," where the brain replays sequences of activity experienced during wakefulness. However, real-world environments are inherently stochastic. For instance, if event A leads to event B 70% of the time and event C 30% of the time, the brain must learn these specific odds. Traditional Hebbian learning—often summarized as "fire together, wire together"—is effective at linking sequences but struggles to capture nuanced transition probabilities. Most existing models either produce rigid, deterministic sequences or fail to maintain the delicate excitatory-inhibitory (E-I) balance required for stable neural replay.

Methods: Dual-Plasticity and Network Architecture

To address these challenges, the authors propose a recurrent spiking neural network (RSNN) governed by a dual-plasticity approach. The network consists of Poisson spiking neurons and utilizes a "division of labor" between excitatory and inhibitory populations.

Overview of the model architecture (b, c) and the comparison between externally driven and internally generated transitions (a).
  1. Prediction-Based Excitatory Plasticity: Instead of simple correlation, excitatory synapses are modified to minimize the "prediction error" between a neuron’s somatic activity (driven by external stimuli) and its dendritic activity (driven by internal recurrent connections). Mathematically, this is equivalent to minimizing the Kullback-Leibler (KL) divergence between the network's internal model and the environment's actual statistics.
  2. Inhibitory Plasticity: Inhibitory synapses are modified to maintain E-I balance. They "predict" the excitatory potential to ensure that the network remains stable, preventing it from either saturating or becoming silent.

Key Results and Biological Validation

The model's performance was validated using both synthetic Markovian sequences and real-world biological data from songbirds.

  • Statistical Fidelity: The transition probabilities of spontaneous activity in the trained model showed a striking 1:1 correlation with the "ground truth" probabilities of the training stimuli. This suggests that the error-correction mechanism allows the network to weight its connections proportionally to the frequency of experienced transitions.
  • Emergent Clustering: When trained on complex graphs with "communities" (clusters of states that frequently lead to one another), the network naturally developed low-dimensional representations mirroring these structures, a phenomenon observed in human psychological studies.
  • Replicating Songbird Neural Dynamics: The model successfully replicated experimental findings from the HVC (High Vocal Center) of Bengalese finches. Specifically, it provided a mechanical explanation for why neural predictability drops when a bird reaches a "choice point" in its song. The model shows that this is not merely noise, but the result of the brain trial-averaging across multiple possible stochastic paths.

Discussion and Future Work

While the model successfully embeds first-order Markovian dynamics, real-world sequences often involve longer-term dependencies. Future work could integrate working memory or short-term plasticity to address these complex temporal structures. Additionally, the current model assumes direct target signals for all excitatory neurons; exploring how "hidden" neurons in hierarchical structures learn remains an open question. Finally, incorporating the diversity of inhibitory interneurons (e.g., PV+ or SOM+ cells) could further refine our understanding of how stability and sequence learning are balanced in the biological brain.

Summary updated

Abstract

The brain learns an internal model of the environment through sensory experiences, which is essential for high-level cognitive processes. Recent studies show that spontaneous activity reflects such a learned internal model. Although computational studies have proposed that Hebbian plasticity can learn the switching dynamics of replayed activities, it is still challenging to learn dynamic spontaneous activity that obeys the statistical properties of sensory experience. Here, we propose a pair of biologically plausible plasticity rules for excitatory and inhibitory synapses in a recurrent spiking neural network model to embed stochastic dynamics in spontaneous activity. The proposed synaptic plasticity rule for excitatory synapses seeks to minimize the discrepancy between stimulus-evoked and internally predicted activity, while inhibitory plasticity maintains the excitatory-inhibitory balance. We show that the spontaneous reactivation of cell assemblies follows the transition statistics of the model's evoked dynamics. We also demonstrate that simulations of our model can replicate recent experimental results of spontaneous activity in songbirds, suggesting that the proposed plasticity rule might underlie the mechanism by which animals learn internal models of the environment.
BibTeX
@inproceedings{asabuki2024embedding,
  author = {Toshitake Asabuki and Claudia Clopath},
  title = {{Embedding stochastic dynamics of the environment in spontaneous activity by prediction-based plasticity}},
  booktitle = {eLife},
  year = {2024},
  doi = {10.7554/elife.95243},
}
Figure