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 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
Neuroscientists have long observed that even when we aren't processing external stimuli, our brains are buzzing with activity. This "spontaneous activity" often mirrors past experiences—a phenomenon known as replay. While we have models for how the brain learns fixed sequences (like a song), it has been much harder to explain how the brain learns stochastic dynamics.
In the real world, if event A happens, event B might follow 90% of the time, while event C follows only 10% of the time. How does a network of neurons learn these specific "transition probabilities" so that its spontaneous "daydreaming" accurately reflects the odds of the real world? Traditional Hebbian learning ("neurons that fire together, wire together") is often too rigid to capture these subtle statistical nuances.
Methods
The authors developed a Recurrent Spiking Neural Network (RSNN) consisting of excitatory (E) and inhibitory (I) neurons. The core innovation lies in a "division of labor" between two new synaptic plasticity rules:
- Excitatory Plasticity (The Predictor): These synapses use an error-based rule. They try to minimize the difference between the actual stimulus-evoked activity and the network's internal prediction of what should happen next. Essentially, the E-synapses are "trained" to become a predictive engine of environmental transitions.
- Inhibitory Plasticity (The Balancer): To prevent the network from becoming over-excited or unstable during this learning process, inhibitory synapses adapt to maintain a precise "E-I balance." They learn to predict the recurrent excitatory potential, acting as a stabilizing force that allows structured spontaneous activity to emerge.
The neurons themselves were modeled as inhomogeneous Poisson processes. The network was tested on its ability to learn Markovian transition structures, complex community structures, and compared against existing experimental recordings from the HVC in Bengalese finches (Bouchard and Brainard, 2016).
Key Insights and Results
- Error-Based vs. Correlation-Based: Moving beyond simple Hebbian correlation, this model shows that predictive error is a powerful signal for embedding complex statistics into brain circuitry.
- The Necessity of Inhibition: A standout finding is that excitatory learning alone isn't enough. Without the adaptive inhibitory "balancing" rule, the spontaneous activity becomes blurred and fails to reflect the learned transition probabilities.
- Statistical Accuracy: After training, the spontaneous transitions between neural assemblies showed a striking 1-to-1 correlation with the transition probabilities of the training stimuli.
- Biological Validation: The model’s "post-stimulus activity" (PSA) matched the statistical signatures found in Bengalese finch HVC recordings, providing a plausible mechanistic explanation for how birds learn their complex songs. Specifically, the model replicates how neural predictability decreases as the uncertainty of upcoming song syllables increases.
Discussion and Limitations
While the study is a significant step forward, several areas for future work remain. The current model focuses on first-order Markov processes; real-world sequences often have longer-range dependencies requiring mechanisms like working memory. Additionally, the plasticity rules are heuristic; future work could seek a more formal theoretical derivation from a strict mathematical objective like Maximum Likelihood. Finally, testing the rules in more biologically diverse architectures—accounting for realistic E/I ratios and multiple types of plastic synapses—is a logical next step.
Summary updated