Contribution of task-irrelevant stimuli to drift of neural representations

Farhad Pashakhanloo

Abstract

In both biological and artificial neural networks, internal representations often undergo continuous change—a phenomenon known as "representational drift"—even when the system's external behavior remains stable. This paper investigates the underlying causes of this drift, demonstrating that it is significantly driven by task-irrelevant stimuli. By treating these irrelevant inputs as a source of "learning noise," the study shows how they push network representations along a manifold of equivalent solutions. Through theoretical modeling and computational simulations, the author establishes that this drift is not merely random noise but a structured byproduct of the learning process itself.

Introduction

A fundamental puzzle in neuroscience and machine learning is why neural representations shift over time. A neuron that responds to a specific stimulus today may become unresponsive to it a month later, while another neuron takes its place, all while the organism's performance on the task remains unchanged. While previous hypotheses attributed this drift to biological "glitches" like synaptic turnover, this research explores whether drift is an inherent consequence of how networks process information.

Specifically, the paper examines the role of task-irrelevant stimuli—information that the network has learned to ignore. Even when a network successfully filters out these distractors to maintain stable performance, the stimuli continue to interact with the learning mechanism, inducing a slow diffusion of weights that manifests as representational drift.

Methods and Theoretical Framework

To analyze the dynamics of drift, the author employs a combination of theoretical physics-inspired modeling and diverse computational simulations.

Stochastic Differential Equations (SDEs)

The study utilizes SDEs to decompose network dynamics into two distinct components:

  1. Normal Fluctuations: Fast changes that keep the network near a functional solution.
  2. Tangential Diffusion: Slow drift that occurs along the "manifold" of equally effective solutions where the task error remains constant.

Learning Paradigms

To ensure the findings are broadly applicable, the author tests several learning rules:

  • Hebbian-based Rules: Including Oja’s rule and Similarity Matching, which are often cited for their biological plausibility.
  • Gradient-based Rules: Stochastic Gradient Descent (SGD) applied to linear autoencoders and supervised two-layer networks.

Analytical Derivations

The paper derives "diffusion coefficients" to quantitatively predict the rate of drift based on the statistical properties of the input data, allowing for a direct comparison between theoretical predictions and experimental simulations.

Experimental Setup and Datasets

The theory was validated across three primary environments:

  • Synthetic Gaussian Data: A controlled environment used to manipulate the variance and dimensionality of task-irrelevant noise.
  • Nonlinear Ring Task: A task where a network represents positions on a circle, allowing for the visualization of how neuronal "receptive fields" move over time.
  • MNIST: The standard handwritten digit dataset, used to confirm that the theoretical findings hold for high-dimensional, real-world data.

Key Insights

The Impact of "Filtered-Out" Noise

The central finding is that stimuli the network has learned to ignore still contribute to weight updates. Because these updates do not affect the final task performance, they act as a "random walk" in the weight space. This suggests that drift is a fundamental byproduct of the stream of data an agent receives.

Anisotropic Drift

Unlike generic synaptic noise, which might push weights randomly in any direction, learning-induced drift is "anisotropic." It follows specific geometric patterns dictated by the structure of the task-irrelevant data. This creates a unique "signature" of drift that depends on the input statistics.

The Dimensionality Trade-off

The research reveals a non-monotonic relationship between the number of neurons (representation dimension) and the drift rate. Initially, increasing the number of neurons provides more "room" for representations to move, increasing drift. However, as the representation dimension approaches the total input dimension, the available "irrelevant space" shrinks, and the drift eventually vanishes.

Results

  • Scaling Laws: The drift rate ($D$) was found to scale with the cube of the learning rate ($\eta^3$) and the square of the noise variance ($\lambda_\perp^2$).
  • Dimensionality Dependency: In Hebbian networks, the drift rate is proportional to $(m-1)(n-m)$, where $n$ is the input dimension and $m$ is the output dimension.
  • Simulation Match: Theoretical predictions showed an excellent match with simulations across all tested architectures, including the complex, non-monotonic behaviors observed in MNIST experiments.
  • Qualitative Signatures: The paper demonstrates that drift caused by task-irrelevant data is qualitatively different from drift caused by simple Gaussian synaptic noise, providing a framework for experimentalists to distinguish between these sources in biological data.

Discussion and Future Directions

This work suggests that by observing the specific patterns of drift in a biological brain, researchers might be able to "reverse-engineer" the underlying learning algorithms the brain employs.

However, the study notes several limitations. The core mathematical derivations are currently limited to linear networks, though simulations suggest they hold for nonlinear systems (like those using ReLU activations). Additionally, the study assumes a stationary environment; in the real world, what is "relevant" often changes, which would likely introduce further complexity to the drift dynamics. Future work could involve testing these predictions in neuroscience experiments by precisely controlling the amount of distractor stimuli presented to an animal during a task.

Summary updated

BibTeX
@inproceedings{pashakhanloo2025contribution,
  author = {Farhad Pashakhanloo},
  title = {{Contribution of task-irrelevant stimuli to drift of neural representations}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=jAoqtT58G4},
}
Figure