🧠 neuroscience

A network model of path control

Inspired by the insect fan-shaped body, this paper presents a recurrent neural network model that uses memory of past movements to dynamically shape an agent's velocity setpoints, thereby enabling the generation of diverse exploration paths and the stabilization of lower-level motor control.

Original authors: Osman, M. A. M., Zhang, W., Pehlevan, C., Wilson, R. I.

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Osman, M. A. M., Zhang, W., Pehlevan, C., Wilson, R. I.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the quiet, dark world of the insect brain, a small structure called the central complex acts as a compass and a map. Deep inside this region lies the fan-shaped body, a part of the brain that helps insects like flies and bees know which way they are facing and how to get home. Scientists have long known that this area contains layers of neurons arranged in columns, like the pages of a book, where activity patterns encode the animal's direction and speed. For years, researchers believed this structure worked like a simple calculator, adding up every step an insect takes to figure out its exact position relative to where it started. This process, known as path integration, is how a foraging ant might return to its nest after a long, winding journey. However, recent observations have shown that the fan-shaped body is far more intricate than a simple calculator, containing many parallel pathways and complex connections that do not always fit the old model of perfect, continuous tracking.

A team of researchers at Harvard University has now built a computer model to understand what this complex network actually does. Instead of assuming the brain simply sums up steps, they proposed that the fan-shaped body acts as a sophisticated filter that shapes how an animal moves through its world. In their model, a virtual agent receives a signal representing its current speed and direction. This signal enters a network of interacting neurons that remember past movements. The network then uses this memory to set a target direction, or a "setpoint," for the agent to follow. If the agent drifts away from this target, the network signals a correction, slowing the agent down and steering it back on course. Crucially, the agent is also influenced by other, simpler systems that react to immediate surroundings, like a sudden gust of wind or a nearby obstacle. The fan-shaped body does not control every tiny movement; instead, it provides a higher-level guide that organizes these lower-level reactions into a coherent path.

The researchers found that by adjusting how the neurons in their model connect to one another, they could produce a wide variety of movement patterns. If the connections are set to remember the recent past for a short time, the agent moves in a straight line but wobbles slightly, exploring its immediate area. If the connections are tuned to remember for longer, the agent moves in a more persistent, straight line, covering more ground efficiently. The model also showed that by including specific types of connections that reverse the phase of the signal, the network could generate rhythmic, back-and-forth movements. This kind of oscillation is useful for a systematic search, allowing an animal to scan a small area thoroughly before moving on. In essence, the network extracts features from the history of movement—such as how long the agent has been traveling in one direction or how often it has turned—and uses that information to decide the overall shape of the journey.

This work suggests that the fan-shaped body is not just a passive recorder of position, but an active shaper of behavior. It allows an insect to balance between two competing needs: the urge to explore new territory quickly and the need to search a specific spot carefully. The model demonstrates that the brain can achieve this by filtering the history of movement through different time scales. For instance, if an insect spots a food source that appears and disappears, the network can help the animal maintain its course toward that source even when the signal is temporarily lost. By remembering the direction of the last sighting, the network stabilizes the movement, preventing the animal from wandering aimlessly during the gaps. This ability to hold onto a goal despite intermittent distractions is a form of memory that organizes behavior over time.

The study also highlights a fundamental difference between how the brain represents information and how it generates action. In this model, the neurons do not explicitly store a map of the world or a list of coordinates. Instead, they simply process the flow of movement, creating activity patterns that naturally lead to specific types of paths. This approach mirrors how other parts of the brain, like the motor cortex, might work, where the focus is on generating the right dynamics for movement rather than calculating a precise destination. The researchers showed that when these neural dynamics are coupled with the physical movement of the body, they create a closed loop where the past influences the future. The agent is not merely reacting to the world in the present moment; it is carrying a memory of where it has been, and that memory subtly steers it toward a future path.

Ultimately, this research offers a new way to think about navigation and decision-making in the insect brain. It suggests that the complex web of connections in the fan-shaped body allows an animal to tailor its search strategy to the environment. Whether the insect needs to cover a large area quickly or hunt for a hidden prize in a small patch, the network can adjust its internal timing to produce the right kind of movement. This flexibility arises not from a central commander issuing orders, but from the natural dynamics of a network that remembers the past to guide the future. The findings provide a concrete explanation for how a relatively simple brain structure can produce the rich, varied, and adaptive behaviors seen in nature, bridging the gap between neural activity and the complex paths animals take through their world.

Technical Summary: A Network Model of Path Control

Problem Statement
The insect fan-shaped body (FB), a structure within the central complex, is known to contain neural arrays where sinusoidal activity patterns encode navigation-relevant vectors. While feedforward models of these arrays can perform arithmetic operations like vector addition, the computational role of the recurrent ensembles remains unclear. Specifically, recent connectomic and imaging data suggest the FB does not perform classical path integration (producing a static position estimate) but rather exhibits transient, leaky integration dynamics. Furthermore, the FB operates in a closed loop: it receives velocity vector inputs and outputs a "setpoint" that influences the animal's heading and speed via motor neurons (PFL2/PFL3). It is currently unclear how these recurrent dynamics, combined with parallel lower-level sensorimotor pathways, shape the spatial and temporal structure of an agent's trajectory.

Methodology
The authors developed a linear recurrent neural network model inspired by the architecture of the insect FB to investigate these dynamics.

Network Architecture: The model consists of $N$ populations of neurons representing columnar cell types. Each population is spatially tiled by columns corresponding to compass directions ( $\theta \in [0, 2\pi)$ ). Neurons are coupled via two types of recurrent connections: phase-aligned (same column) and phase-inverted (opposing columns, separated by $\pi$ ). All connections are excitatory, but phase-inverted connections effectively act as negative couplings on the first harmonic of the activity.
Input Encoding: The network receives an external input encoding the agent's velocity vector $v(t)$ as a sinusoidal spatial profile (phasor representation).
Dynamics: The network is modeled as a system of leaky integrators. The activity of each columnar cell type is decomposed into a mean and a first harmonic (sinusoidal component). The authors focus on the first harmonic, which evolves according to linear differential equations driven by the recurrent connectivity matrix ( $W^+ - W^-$ ) and the velocity input.
Closed-Loop Control: The network's output is a setpoint vector $u(t)$ , derived from a weighted sum of the network's activity profiles. This setpoint drives a control policy that continuously corrects the agent's velocity vector toward the setpoint. The agent's velocity is also influenced by a "forcing function" $f(t)$ , representing independent lower-level sensorimotor pathways and environmental stimuli.
Analysis: The system is analyzed using three frameworks: (1) as an autoregressive process where velocity depends on its own history via a temporal filter; (2) as a joint dynamical system of neural activity and velocity; and (3) as a filtered version of the external forcing function.

Key Contributions and Results

Generalization of Path Integration: The model demonstrates that the FB network acts as a bank of temporal filters applied to the velocity input. Depending on the recurrent connectivity (strength of phase-aligned vs. phase-inverted connections), the network can extract trajectory features over various time scales and frequency bands. This generalizes classical path integration (which corresponds to a constant filter) to include leaky integration, double-exponential filtering, and band-pass oscillatory filtering.
Path Structure Control: By tuning the network's coupling parameters, the model can generate diverse path structures suitable for different exploration strategies:
- Persistent Movement: Achieved by strong phase-aligned coupling, allowing the agent to maintain a straight course over long time scales while exhibiting diffusive behavior over short scales.
- Oscillatory/Backtracking: Achieved via phase-inverted connections, generating trajectories with regular reversals useful for local systematic searches.
- Correlated Random Walks: The resulting paths are mathematically equivalent to correlated random walks, where the correlation structure is determined by the network's internal dynamics.
Stabilization of Intermittent Cues: The model shows how the FB can stabilize movement toward a cue that is only intermittently available (e.g., gusts of wind or fleeting visual landmarks). By integrating the forcing function over a longer time scale (via strong recurrent connections or high control loop gain), the network allows the agent to maintain a bias toward the cue direction even when the signal is absent, effectively bridging gaps in sensory input.
Separation of Representation and Generation: Unlike previous models where specific cell types encode specific variables (like position), this model posits that cell types act as linear filters on velocity. The "representation" emerges from the filtering process to generate behavior, rather than explicitly encoding an environmental state.

Significance and Claims
The paper claims that this network model provides a mechanistic explanation for how the FB uses memory of past movement to organize structured behavioral sequences. The authors propose that the FB does not merely compute a position estimate but extracts and stores features of past velocity to generate a path with a desired spatial structure.

Key implications highlighted in the discussion include:

Cognitive Rudiments: The FB exhibits rudiments of cognition, including working memory, evidence accumulation, and the ability to sustain actions initiated by lower-level systems despite distractors.
Embodied Cognition: The model illustrates how behavior and neural dynamics co-evolve in a closed loop. The network "observes" the agent's behavior (via velocity input) and adjusts its setpoint accordingly, rather than issuing abstract commands to a passive body.
Hierarchical Control: The model suggests a control architecture where lower-level pathways (represented by the forcing function) act as the primary engine of behavior, while the high-level recurrent network modulates the persistence and time scales of these behaviors. This contrasts with traditional views of high-level brain regions as the sole initiators of action.

The authors emphasize that these results are derived from analyzing observed connectivity patterns and do not rely on training the network to perform a specific task, offering a direct link between neural circuit structure and emergent behavioral dynamics.

Technical Summary: A Network Model of Path Control

More like this