Network Model of State-Dependent Sequencing

Jeffrey P. Sutton, Adam N. Mamelak, J. Allan HobsonNeurIPS 1991

This paper, published in the early 1990s, represents a fascinating bridge between computational neuroscience and the study of consciousness. It explores how the chemical "weather" of the brain—neuromodulation—fundamentally changes how we process sequences of information during wakefulness and sleep.

The Challenge of Dynamic Brain States

Most artificial neural networks are designed with a single, static mode of operation. However, biological brains are highly dynamic, undergoing massive shifts in state during the sleep-wake cycle.

The challenge the authors tackle is explaining how the same physical network of neurons can support vastly different "styles" of information processing. Why is our waking thought logical and progressive, while our dreams (REM sleep) are characterized by "bizarreness," incongruous memories, and sudden jumps in narrative?

State-dependent sequencing mapping

The authors propose that these cognitive shifts aren't due to changes in the network's hardwiring, but rather changes in the chemical environment (neuromodulation) that alter how neurons respond to signals. By adjusting parameters representing aminergic and cholinergic systems, their model replicates the transition from stable, linear sequencing (Wake) to the associative, bizarre jumps characteristic of REM sleep.

Modeling Neuromodulation

To test this, the authors developed an associative sequencing network consisting of 50 binary neural elements. The model's uniqueness lies in its two-part synaptic architecture and its modulatory "knobs":

  1. Dual Connectivity:
  2. One set of connections encodes static memory patterns via Hebbian learning.
  3. A second set of asymmetric connections encodes the transitions between these patterns, creating temporal loops (e.g., Memory A1 leads to A2, which leads to A3).
  4. Neuromodulatory Parameters:
  5. $\beta$ (Aminergic Modulation): Modeled as a gain factor or "inverse temperature." High $\beta$ (Wake) increases the signal-to-noise ratio, making the network's behavior deterministic and high-fidelity. Low $\beta$ (Sleep) makes the network noisy and unstable.
  6. $\delta$ (Cholinergic Modulation): Modeled as phasic excitatory bursts, representing "PGO waves" (ponto-geniculo-occipital) that occur during REM sleep. These bursts provide a sudden "kick" to the membrane potential of neurons.

The network uses a parallel updating rule where the probability of a neuron firing depends on its input scaled by $\beta$ and $\delta$. The authors used synthetic, uncorrelated memory patterns generated at random to test the network's dynamics.

Key Insights

  • Modulation as a Filter, Not a Carrier: Neuromodulators like norepinephrine or acetylcholine don't carry the "content" of our thoughts. Instead, they act as filters that amplify or dampen the distributed information already embedded in the network's synapses.
  • Bizarreness as Bifurcation: In the simulated REM state (low $\beta$, high $\delta$), the network experiences "bifurcations." A PGO-like burst can knock the network out of one memory loop (Loop A) and into another (Loop B). This provides a computational explanation for "dream bizarreness"—the sudden, illogical shifts in a dream's setting or characters.
  • The "Defibrillation" Effect: Interestingly, the authors found that in a state of high disorder, a phasic cholinergic burst could sometimes act as a "defibrillator," pushing a chaotic network back into an orderly, sequential progression.
  • State-Dependent Fidelity: The "fidelity" of our cognitive sequencing is chemically dependent. As aminergic levels drop (moving from Wake to NREM to REM), the network becomes increasingly sensitive to small perturbations, leading to ruminative or perseverative patterns.

Simulated States and Results

By tweaking the $\beta$ and $\delta$ parameters, the researchers successfully simulated three distinct cognitive states:

  • Wake State ($\beta=2.0, \delta=0$): The network showed high fidelity, staying strictly within its assigned memory loop until an external input forced a change.
Simulated Wake State
  • NREM State ($\beta=1.1, \delta=0$): Sequencing persisted, but with significant fluctuations. The network was "ruminative"—it stayed in the loop but was far less stable than in the wake state.
  • REM State ($\beta=1.0, \delta=2.5$ bursts): The network produced "bizarre" outputs. It generated mixed memories (e.g., a pattern that was a hybrid of two different memories) and experienced sudden discontinuities, jumping between unrelated sequences. The authors noted that lower-intensity, longer-duration bursts were more effective at inducing these "dream-like" shifts than short, high-intensity spikes.

Limitations and Future Work

While groundbreaking, the model has constraints. It uses simple binary neurons ($+1/-1$), which lack the complexity of real biological neurons like spiking rates and refractory periods. Furthermore, with only 50 elements, the model serves as a proof of concept rather than a full-scale simulation of a brain region.

Crucially, the synaptic weights in this study were fixed before the simulation. The authors suggest that the next frontier is state-dependent learning, where neuromodulators don't just change how the network runs, but also how it learns (plasticity) in real-time. This future direction could finally explain why we struggle so profoundly to form long-term memories of our dreams.

Summary updated

BibTeX
@inproceedings{sutton1991network,
  author = {Jeffrey P. Sutton and Adam N. Mamelak and J. Allan Hobson},
  title = {{Network Model of State-Dependent Sequencing}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1991},
}
Figure