Abstract
In recent years, there have been many computational simulations of spontaneous neural dynamics. Here, we explore a model of spontaneous neural dynamics and allow it to control a virtual agent moving in a simple environment. This setup generates interesting brain-environment feedback interactions that rapidly destabilize neural and behavioral dynamics and suggest the need for homeostatic mechanisms. We investigate roles for both local homeostatic plasticity (local inhibition adjusting over time to balance excitatory input) as well as macroscopic task negative activity (that compensates for task positive, sensory input) in regulating both neural activity and resulting behavior (trajectories through the environment). Our results suggest complementary functional roles for both local homeostatic plasticity and balanced activity across brain regions in maintaining neural and behavioral dynamics. These findings suggest important functional roles for homeostatic systems in maintaining neural and behavioral dynamics and suggest a novel functional role for frequently reported macroscopic task-negative patterns of activity (e.g., the default mode network).
Introduction
Most computational neuroscience models simulate the brain "at rest" or during very constrained, one-way tasks, such as processing a single static image. However, biological brains exist in a continuous closed-loop: neural activity drives movement, movement changes sensory input, and sensory input further alters neural activity.
This paper demonstrates that in a simple embodied simulation, this feedback loop is inherently unstable. Without regulation, an agent might get "trapped" against a wall because constant sensory input keeps motor neurons firing into it, or the brain might go silent in empty space. The challenge is finding a way to keep the brain in a "Goldilocks zone" of activity—neither saturated nor silent—despite the unpredictable nature of the environment.
Methods
The researchers utilized a three-layered approach to build their simulation:
- The "Brain" Structure: They used a structural connectivity matrix derived from human Diffusion Tensor Imaging (DTI), mapping connections between 66 cortical regions (Hagmann et al., 2008).
- Neural Dynamics: They employed the Greenberg-Hastings model, where each node can be in one of three states: Excitatory, Quiescent, or Refractory. A node fires if it receives enough input from its neighbors or via a small random probability.
- The Balancing Acts:
- Local Homeostasis: Each node has an adjustable "threshold." If a node fires too much, the threshold rises (increasing inhibition); if it fires too little, the threshold drops.
- Macroscopic Task-Negative (TN) System: Inspired by the Default Mode Network (DMN), specific nodes were designated to be "anti-correlated" with sensory input. When sensory nodes are active, TN nodes are forced off, and vice versa, acting as a global "counterweight" to external stimulation.
The virtual agent was placed in a 2D box with visual and collision sensors. Its movement was controlled by the activity of specific "motor" nodes in the network model.
Key Insights: The DMN as a "Keel"
A striking insight from this work is the functional interpretation of the Default Mode Network. Rather than just being for internal thought or "daydreaming," the authors suggest it acts like the keel of a sailboat. It doesn't move the boat forward, but it counterbalances the forces on the sail, providing the stability needed for the boat to handle a wider range of environmental "winds" (sensory inputs).
Furthermore, the simulation was most stable when the Task-Positive (sensory) nodes and Task-Negative nodes were structurally close to each other in the network. This suggests that the physical wiring of the brain may be optimized to facilitate this balancing act.
Results
Both local and macroscopic mechanisms successfully kept the average neural activity near the target rate.
Reduced Environmental Sensitivity
In the model with the TN system, the correlation between "distance from the wall" and "neural activity" dropped significantly (Spearman Rho moved from -0.45 to -0.25). This indicates that the brain's internal state became less "hijacked" by its immediate proximity to obstacles.
Enhanced Exploration
Agents with the TN system showed higher fractal dimension and entropy in their movement trajectories. Qualitatively, they did not just get stuck in corners; they traversed the environment in more complex, varied patterns.
All reported differences between the "local-only" and "local + TN" models were highly significant ($t > 3$).
Discussion
While this is a "toy model" where nodes do not perform complex computations like object recognition, it highlights the critical role of homeostatic systems in maintaining stable behavioral dynamics. Future work could involve more complex environments and sophisticated neural models that allow for actual learning. The researchers also suggest that exploring how "lesioning" the task-negative system in the model might mimic the behavioral instabilities seen in certain neurological or psychiatric disorders.
Summary updated