A neural network model of sensoritopic maps with predictive short-term memory properties.

J Droulez, A BerthozProceedings of the National Academy of Sciences 1991 · DOI

This 1991 paper by Droulez and Berthoz presents a landmark neural network model explaining how our brains maintain a stable "map" of the world despite our eyes constantly moving. By using internal feedback from motor commands to shift activity within a retinotopic map, the authors demonstrate how "egocentric" coordinates can achieve the same goal-invariance as a fixed world-coordinate system.

TL;DR

The authors propose a neural network that maintains and updates the memory of a target's position by "shifting" a peak of neural activity in response to eye-velocity signals. This model proves that the brain doesn't need a static, world-fixed coordinate system to track objects; it can simply update its sensor-based maps using internal copies of motor commands.

What Problem Does This Paper Address?

When you look around, your eyes perform rapid jumps called saccades. If a target disappears and your eyes move before you can look at it, you can still accurately saccade to where that target was. This is "goal invariance."

Historically, researchers proposed two ways the brain handles this:

  1. Allocentric (World-fixed): The brain converts the target's position into a stable, room-centered coordinate system.
  2. Egocentric (Sensor-fixed): The brain keeps the target in retinal coordinates but continuously "re-maps" it based on how the eyes move.

The problem was that while the brain's anatomy (like the superior colliculus) is clearly organized in retinotopic (egocentric) maps, there was no clear mechanism for how these maps could "update" themselves without losing the target's location during a movement.

Methods

The authors developed a formal neural network consisting of three layers of neurons organized into a topologically ordered map:

  • Input Neurons (I): These receive visual signals and compare them to internal predictions.
  • Main Neurons (M): These act as short-term memory, holding a "mountain" of activity representing the target.
  • Multiplier Interneurons (P): These combine the activity from the M-layer with an efferent copy (an internal copy of the eye-velocity signal).
Neuronal architecture of the dynamic memory map

The core mathematical trick is a first-order Taylor expansion. To shift a function $f(x)$ by a small amount $Dx$, you can approximate the new state using the current state and its spatial gradient: $f(x + Dx) \approx f(x) + Dx \cdot f'(x)$. In this model, the P-neurons calculate this gradient-based shift, effectively "sliding" the activity peak across the map in the opposite direction of the eye movement.

Datasets

The study used simulated data rather than biological datasets. The authors performed quantitative simulations on:

  • A 1D map consisting of 50 modules.
  • A 2D map consisting of a 31x31 grid of modules.
  • Visual inputs were modeled as Gaussian-like functions.
  • The network was "trained" using a modified Hebbian learning rule where the intermittent visual signal served as the "teacher" to help the network learn the correct synaptic weights for prediction.

Key Insights and Ideas

  • Stability through Nonlinearity: The "memory" of the target is maintained by the nonlinear, sigmoid-shaped activation of the M-neurons. This prevents the activity from diffusing or disappearing in the absence of visual input.
  • Predictive Remapping: Instead of reacting to sensory change, the network predicts the change. The shift happens because the motor command (eye velocity) is fed directly into the map.
  • Spatial vs. Frequency Coding: The model uses a spatial code (where the activity is matters more than how fast neurons fire). This allows the brain to track multiple targets simultaneously, though it predicts that if targets are too close, their activity "mountains" will merge into a single centroid.
  • Biological Plausibility: The model aligns with observed neural behavior in the posterior parietal cortex, where neurons fire based on a combination of retinal stimulation and eye position.

Key Results

  • Long-term Stability: In the 1D simulation, the stored activity remained stable for at least 50 seconds after the target disappeared.
  • Accurate Shifting: The 2D simulation showed that distances between multiple target peaks were preserved even after oblique saccades.
  • Optimal Velocity: The prediction error was minimized at a specific saccadic velocity (~230 pixels/sec), suggesting the network's connectivity is tuned for natural movement speeds.
Stability of stored activity in the map over time

Limitations and Future Work

  • Precision Trade-off: The authors acknowledge that spatial coding is less precise than frequency coding. The accuracy depends heavily on the "overlapping factor" (how much receptive fields overlap).
  • Velocity Signal Source: The model assumes a clean eye-velocity signal is available. In reality, it's unclear if the brain uses actual velocity, expected velocity, or a reconstruction from successive positions.
  • Generalization: While focused on eye movements, the authors suggest this "predictive remapping" could apply to head movements, body orientation, and even limb control in reaching tasks. Future work would involve integrating vestibular (inner ear) signals to update maps during head rotation.

Summary updated

Abstract

Coordinated orienting movements can be accurately performed without direct sensory control. Ocular saccades, for instance, have been shown to be reprogrammed after target disappearance when an intervening eye movement is electrically triggered before the saccade onset. Saccadic eye movements can also be executed toward memorized targets, even when the subject has been passively moved in darkness. Two hypotheses have been proposed to account for this goal-invariance property: either (i) the goal is reconstructed and memorized in the stable frame of reference linked to the environment ("allocentric, coordinates") or (ii) the goal is selected and memorized in the sensors-related maps ("egocentric coordinates") and is continuously updated by efferent copies of the motor commands. In this paper, we shall describe a formal neural network based on this second hypothesis. The results of the simulation show that target position can be memorized and accurately updated in a topologically ordered map, using a velocity-signal feedback. Moreover, this network has been submitted to a simple learning procedure by using the intermittent visual recurring afferent signal as the teaching signal. A similar mechanism could be involved in control of limb movement.
BibTeX
@inproceedings{droulez1991neural,
  author = {J Droulez and A Berthoz},
  title = {{A neural network model of sensoritopic maps with predictive short-term memory properties.}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {1991},
  doi = {10.1073/pnas.88.21.9653},
}
Figure