Modeling the contribution of theta-gamma coupling to sequential memory, imagination, and dreaming

How does the brain organize a lifetime of experiences into a coherent timeline? For decades, neuroscientists have observed a rhythmic dance in the hippocampus—the brain's memory hub. Fast gamma oscillations (~40 Hz) are nested within slower theta rhythms (~4 Hz). This "theta-gamma code" is believed to be the mechanism that allows us to pack individual moments into a sequential story.

While the existence of this code is well-documented, building a computational model that captures how it moves from simple memory storage to complex states like dreaming or even pathological delusions has proven difficult. Most existing models struggle to explain how the brain transitions between these states. They also struggle to model how an external pacemaker, like the medial septum, regulates the rhythm. A new study from the University of Bologna addresses these gaps. The authors propose a three-layer neural mass model (a mathematical description of population-level dynamics) that links rhythmic coupling directly to different mental states.

Moving beyond internal rhythm generation

Current neurocomputational approaches often treat brain rhythms as emergent properties of the network itself. This means the rhythms only appear after the network has been trained to remember something. While mathematically convenient, this approach fails to capture the biological reality of the hippocampus. In living organisms, the theta rhythm is often driven by external structures, such as the medial septum. It can also be suppressed or modulated rapidly during shifts in consciousness, such as the transition from wakefulness to sleep.

The authors argue that previous models cannot easily simulate the sudden disappearance of theta waves during quiet wakefulness. They also cannot easily model how pharmacological interventions can alter these rhythms. This limitation makes it difficult to use such models to study neurological disorders where rhythmic synchronization is disrupted. To solve this, the researchers shifted the paradigm. Instead of letting the network "invent" its own rhythms through training, they introduced an external "Theta generator" to act as a rhythmic pacemaker. This allows for much greater flexibility in simulating different physiological conditions.

A three-layer architecture for episodic flow

The researchers implemented a "neural mass model." This is a mathematical framework that describes the collective activity of large populations of neurons rather than simulating every single cell. This provides a middle ground between simple models and computationally expensive single-neuron simulations. The architecture consists of three distinct layers designed to mimic the hippocampal circuit:

1. The WM (Working Memory) Layer: Simulating the prefrontal cortex, this layer maintains a stable signal (a "cue") using an auto-excitatory loop. This ensures information isn't lost immediately after a stimulus disappears.
2. The L1 (Layer 1) Layer: Mimicking the CA3 region, this layer functions as an auto-associative memory. It uses Hebbian training (strengthening connections between active neurons) to ensure that if you remember one part of an event, you can recover the rest.
3. The L2 (Layer 2) Layer: Representing the CA1 region, this layer forms a hetero-associative link with L1. This allows the network to transition from one completed episode to the next in a sequence.

Crucially, the model relies on three types of synapses to manage the "traffic" of information. Excitatory synapses ($W_p$) recover features within an episode. Inhibitory synapses ($W_f$) prevent runaway excitation. The most vital addition is the "desynchronizing" synapse ($A_f$). This uses anti-Hebbian learning (rules that weaken connections between simultaneously active neurons) to ensure that features from different episodes do not overlap or interfere with one another .

This prevents the "blurring" of separate memories.

Quantifying the theta-gamma code

The authors tested the computational simulation's ability to retrieve sequences of episodes. They used "orthogonal" sequences (where episodes share no common features) and "non-orthogonal" sequences (where episodes share some features). Under basal conditions, the paper reports high success rates for sequence retrieval. For orthogonal sequences, the model successfully recovered the first episode in approximately 70% of theta cycles. It maintained near-perfect accuracy for the middle episodes. Accuracy dropped slightly to roughly 89% by the fifth episode [Table 1].

The researchers also performed a sensitivity analysis on frequency. They found a strict trade-off between the speed of the rhythms and the length of the memory. Decreasing the theta frequency (e.g., to 2.5 Hz) improves the recovery of longer sequences. This provides more time within a single cycle to process information. Conversely, increasing the gamma frequency to 45 Hz significantly boosts accuracy. It reaches over 95% for all five episodes. However, this introduces a new risk: the emergence of "spurious" episodes. In this case, the network accidentally pulls in fragments from unrelated memories [Figure 8B].

From dreaming to delusion

The most striking aspect of the study is how varying the "noise" and synaptic strength can simulate diverse mental states. By isolating the L1 layer from external input and increasing random noise, the authors simulated "imagination." In this state, the network randomly replays stored sequences.

By further increasing the noise and reducing synaptic strength by two-thirds, the model enters a "dreaming" state. In this mode, the network no longer retrieves perfect sequences. Instead, it creatively recombines shared features from different episodes to generate entirely new, "oneiric" combinations [Figure 9B]. This mimics the fragmented, associative nature of REM sleep.

Finally, the authors modeled a "schizophrenia" state. They did this by reducing the strength of the fast inhibitory desynchronizing synapses. The result is a breakdown of temporal boundaries. The network begins to superimpose features from different episodes onto one another. This effectively simulates a "delusional" state where distinct realities are erroneously merged [Figure 9E].

Assessing the hippocampal blueprint

This model is a powerful proof-of-concept. However, it is a computational simulation rather than a biological subject. The authors admit several key simplifications. The model assumes a "buffer" mechanism to pass episodes between layers. This mechanism currently lacks direct physiological evidence. Additionally, the model keeps the representation of features constant across layers. This ignores the hierarchical abstraction seen in real cortical processing.

Is this model ready for clinical application? Not yet. It remains a theoretical tool. It is designed to help researchers identify which specific synaptic mechanisms might be failing in patients. For example, it helps investigate how disruptions in GABAergic inhibition relate to schizophrenia. By treating theta and gamma rhythms as independent, controllable variables, the authors have provided a flexible framework. We may eventually use such tools to understand how transient hippocampal memories are consolidated into permanent knowledge.

Figures from the paper